Is Cantor Infinity in Mathematic Necessary? Dedicated to the memory of my friend and colleague Jean van Heijenoort Infinity is up on trial (Bob Dylan, Visions of Johanna) Introduction Since the rise of abstract mathematics in Greek times, mathematicians have had to grapple with the problems of infinity in many guises. When ‘mathematics became an integral part of physical science it could be used to formulate precise answers to age-old questions: Is space infinite? Did time have a beginning? Will it have an end? Modern cosmological theo- ries now marshal considerable physical evidence to support the finitude of space and time. But whether or not (or how) infinity is manifested in the physical universe, mathematics requites at its base the use of various in- finite arithmetical and geometrical structures. Without these no coherent system of mathematics is possible, and since mathematics is essential for the formulation of physical theories, there is also no science without these uses of the infinite at the base. Beginning in the 1870s, Georg Cantor came to realize that one must distinguish different onders or sizes of infinity of the underlying sets of ob- jects in these basic mathematical structures. As he continued to work out nfnity in mathematica Te Cantor necessary?” first appeared in G. Toraldo di Francia (ed), L'ofinito nella acienca (Infinity 1 Science), Istituto della Baciclopedi= Tealiana, Rome (1987), 151-209 (Fefermian 1987s). It in reprinted here with the kind permission af the publisher, and appears in this volume in two parts, The present Chapter constitutes the first (approximately) two thitds of the original article the last third, which is more technical, appears ax chapter 12, Some nninor corrections and Additions to bring it up-to-date have been made 8Infinity in mathematics 29 the implications of his ideas, he was led to the introduction of a series of transfinite cardinal numbers for measuring these sizes. Many mathe- maticians of Cantor's time were disturbed by his work, partly due to the novelty of his concepts and partly due to the uncertain grounds on which his computations and arguments with the scale of cardinals rested, But some reacted in direct opposition to Cantor's work, for his reintroduction of the “actual infinite” into mathematics (ironically, after that seemed to have been finally eliminated from analysis by the previous foundational work of the nineteenth century). One of Cantor's most vigorous and severe critics was his former teacher Leopold Kronecker, who would admit only “potentially infinite” sets to mathematics and, indeed, only those reducible to the natural number sequence 0, 1, 2, 3, Worries about the Cantorian approach were compounded when, around the turn of the century, paradozes appeared in the theory of sets by taking its ideas to what appeared to be their logical conclusion. The most famous ‘and simplest of these contradictions was due to Bertrand Russell, but ear- lier ones had already been discovered for Cantor's transfinite numbers by Cesare Burali-Forti? and even by Cantor himself. While these paradoxes did not worry Cantor, the vague distinctions between the transfinite and the “absolute” infinite that he made in order to avoid them could not at first be made precise. But the contradiction plagued Russell and he at- tempted many solutions to escape them: that is, he sought precise system- atic grounds for accepting substantial portions of Cantor's theory while excluding the paradoxes. The means at which Russell finally arrived is, called the theory of types, and though it proved to be very cumbersome as a framework for set theory, it did restore a measure of confidence in Can- tor’s work. Independently, Ernst Zermelo introduced an ariomatic theory of sets, which featured a simple device for limiting the size of sets so as to avoid the paradoxes while providing a very flexible and ready means for the precise development of a considerable portion of Cantor’s theory of higher infinities. Extensions of Zermelo’s axiom system, which are described be- low, allow one to develop Cantorian theory in full while avoiding all known paradoxical constructions. Parallel to the work by logicians providing an axiomatic foundation for higher set theory, Cantor's ideas were put to use more and more in mathe- ‘matics, 50 that these days they are largely taken for granted and permeate the whole of its fabric. But there are still a number of thinkers on the subject who, in continuation of Kronecker’s attack, object to the panoply of transfinite set theory in mathematics, The reasons for doing so are no longer the paradoxes, which have apparently been blocked in an effective ‘way by means of the axiomatic theories devised by Zermelo and his succes- Actually, Bural-Fort's paradex was only implicitly contained in his work. Inciden tally, Russell's paradox was found independently by Zeemelo in 1902. While the work was aot published, Zermelo claimed this earlier discovery in bis 1908 paper, that has Subsequently been confirmed by a variety of evidence in Rang and ‘Thomas (1981)30 Foundational problems sors. Rather, the objections to the Cantorian ideas reside in fundamentally differing views as to the nature of mathematics and the objects (numbers, points, sets, functions, ...) with which it deals. In particular, these oppos- ing points of view reject the assumption of the actual infinite (at least in its nondenumerable forms). Following this up, alternative schemes for the foundations of mathematics have been pursued with the aim to demonstrate that everyday mathematics can be accounted for in a direct and straight- forward way on philosophically acceptable non-Cantorian grounds. While genuine progress has been made along these lines, the successes obtained are not widely known and the alternative approaches have attracted relatively few adherents among working mathematicians. ‘The general impression is that non-Cantorian mathematics is too restrictive for the needs of math- ematical practice, regardless of the merits of the guiding non-Cantorian philosophies, Some logicians would now go farther to bolster this impression by giving. it a theoretical underpinning; their aim is to demonstrate that Cantor's higher infinities are in fact necessary for mathematics, even for its most finitary parts. The results that have been obtained in this direction do, at first sight, appear to justify such a reading. Nevertheless, it is argued below (now in chapter 12] that the necessary use of higher set theory in the mathematics of the finite has yet to be established. Furthermore, a case can be made that higher set theory is dispensable in sctentifically applicable ‘mathematics, that is, in that part of everyday mathematics which finds its applications in the other sciences. Put in other terms: the actual infinite is not required for the mathematics of the physical world. The reasons for this depend on other recent developments in mathematical logic, the description of which is the final aim of this essay [now chapter 12; ef. also chapter 14]. In order to explain the objections to Cantor's ideas in mathematics that lead one to search for viable alternatives, one must first provide some understanding of their nature and use. This will be done here more or less historically though necessarily in outline; we start with a rather innocent looking problem about the existence of certain special kinds of numbers? From Transcendental Numbers to Transfinite Numbers ‘There are two basic number systems (having ancient origins), the set N of natural numbers 0,1,2,3,-.. used for counting, and the set R of real numbers use for measuring. The latter represent positions of points om a ‘two-way infinite straight line, relative to any initial point 0 (the “origin”) and any unit of measurement 1, R is pictured as follows: TPhis is not where Cantor himself started, though he came to it soon enough. For 4 good detailed introduction to the history of the development of Cantor's ideas, see Grattan-Guinness (1980), chepters 5, 6 (by J. W. Dauben and R. Bunn, resp.). For a deeper pursuit would recommend most highly the books of Moore (1982) and Hallett (1984). [CE also Lavine (1994) ]Infinity in mathematics 3u N can thus be identified with a subset of R. Other subsets of use are Z= {... ,—3,—2,-1,0,1,2,3,-..} (the set ofall integers) and Q = {n/m n,m € Zand m # 0}, the set of rational numbers, consisting of all quotients ‘or ratios of integers with nonzero denominators. We here write “x € S” for the relation of membership of an object x to a set S, and “{z: P(z)}” for the set of all objects x satislying a determinate property P(z). If x does not belong to S we write “xg S.” Finite sets $ can be denoted directly by a listing of their elements as $= (00,01,@2,-.. ,@,}. This notation is extended to certain infinite sets such as N= {0,1,2,...,n,-..} and 2 (as above), when we have a complete survey of their elements, If S; and S are sets, then $y is a subset of Sz, in symbols Si C Sz, if for every 1 € Sy we have x € $2; ~ Si = {n: 1 € Sy and ¢ ¢ $;} then denotes the difference of these two sets. $2 —S; might be empty, when Si = Sa; the empty set is denoted by 0. The set Q is densely dispersed throughout R and cannot be distin- guished from R by a simple picture as above. A basic realization from Greek geometry was the existence of irrational magnitudes, that is, ele- ments of R — Q. For, Pythagoras’ law giving the hypotenuse c of a right triangle in terms of its legs a, by ¢? = a+ 6?, or equivalently c = va" +6, leads directly to quantities such as V2 = YI? + 1? and V5 = VI7+%, which are easily proved to be irrational. Other kinds of irrational numbers also arise naturally in geometry, for example, {3 in the classical problem of the duplication of the cube (that is, construction of a cube with double the volume of a given one), and r(= 3.14159...), the ratio of the cireum- ference of a citcle to its diameter. However, the proofs of the irrationality of numbers like + only came much later. Other irrational numbers arise from the solution of algebraic equations; for example, one solution of 24 —7 = Ois 2 = YT and of 28 ~2!—1=0is 4+ V5)/2. Some equations, such as 2? +1 = 0, have no solutions in real numbers, although we can treat their solutions in the extension of the real number system by the imaginary numbers such as Y=T; however, those will not concer us here. A real number is called algebraic if it is the solution x of a nontrivial polynomial equation p(x) = 0 with integer coefficients; that is, p(x) = ky2" + kyi2""! +... +hy2 + ko with n > 0 and k, #0, and cach k, € Z. We use “A” to denote the set of all (real) algebraic numbers. Thus QC A and A contains all the irrational-mumbers shown above, except possibly *. However, it is natural to suspect that A, since there is no known polynomial equation of which itis a root ‘This was in fact first conjectured by Legendre in the eighteenth century, but a proof did not come until a century later.32 Foundational problems A number is called transcendental if it is in the set T = R—A. An other specific number which, like x, is ubiquitous in mathematics and was, conjectured to be transcendental is the base ¢ (= 2.71828...) of “natural” logarithms. The first proof that there erist any transcendental numbers at all, that is, that T 4 @, was given by Liouville in 1844. His method was to find a property P(z) which applies to all algebraic numbers x and which says something (technical) about how well x can be approximated by rational numbers. Liouville then cooked up a real number é which does not satisfy the property P, so € must be transcendental. Infinitely many other transcendental numbers can also be produced in this way, but Li- ouville's method did not help answer the specific questions as to whether € and © are transcendental. ‘Those results were not obtained until some- what later, by Hermite (in 1873) and Lindemann (in 1882), respectively, using rather special methods. In the meantime, Cantor published in 1874 a new and extremely simple but striking argument to prove the existence of transcendental numbers. Cantor's result in this respect. was no better than Liouville's, but the methodology of his proof turned out to be one of the ‘main starting points for his novel conception of infinity in mathematics First of all, Cantor defined two sets Sy and S2 to be equinumerous if their elements can be placed in one-to-one correspondence with each other; symbolically this is indicated by Si ~ Sp. A set $ is finite if it is equivalent to some initial segment of N, possibly empty, that is, S ~ (0, a, where n> 0. A set $ is called denumerable if $ is finite or S~ N. Every nonempty denumerable set $ can be listed a8 S = {a9,@1,.-. :duy---}y possibly with repetitions, and conversely. From this follows directly several basic facts: (1) a denumerable union of denumerable sets is denumerable; (2) the set Q is denumerable; (3) the set A is denumerable. In (1) we are considering sets So, 51,-.-,Sns.-. , each of which is de- numerable and may be assumed nonempty; these are listed as $= (a0, a1, amy} Si= (0, br, ae i Sa= {o, ef eo}Infinity in mathematics 33 ‘The arrows have been added to show that the union S can be listed fol- lowing the indicated arrows as $= {o,.01b0,.425b1)€05+-+} Now, for (2), take Sq to be the set of all multiples m/n for n #0 and ‘m in Z; each S, is denumerable and their union is Q, so (2) follows. To prove (3), one shows first of all that all equations of the form haz” +... kuz + ko = 0 with integer coefficients and km #0 can be enumerated. If a(x) = 0 is the n™ such equation, take S, to be the set of its solutions. ‘This set is finite (possibly empty), hence certainly denumerable. But the set A of algebraic numbers is the union of these S,'s, s0 it also is denumerable; that is, (3) holds. Now, in contrast, Cantor showed that (4) R is non-denumerable, In other words there is no way to list Ras a simple sequence of real num- bers, (a, 21,22 ...}- Cantor's first proof of (4) made use of special prop- erties of R. Later he gave a simpler proof that could be generalized to other sets; this used his famous diagonal argument. It is sufficient to show that the set of reals x between 0 and 1 cannot be enumerated. Indeed, given any enumeration {29,.21,2...} of a subset of R, we shall construct a number x that is not in the enumeration. First write each member of {20,11,12...) as an infinite decimal 1 = O.aya303 0, 21 = O.bibabs By 2 = 0.016208. Now form x = Otukaky...ky... not in (t0,21,22-..} by choosing ky # ayshy # bajks # cp, etc. For example, take ky = if a, ¢ 0 and ky = 1 if a, = 0, etc. This proves (4); it then follows immediately from (3) that A # R. Hence T = R— A must be nonempty, and thos the existence of transcendentals is newly established. In fact, T must also be nondenumerable, for otherwise by (1) and (3) we would have denumerable. ‘This is a stronger conclusion than Liouvill’s, which was merely that T is infinite Clearly there are two senses of the size of a set involved here. If $y is 2 proper subset of Sp, then it is smaller than S in the sense of containing fewer elements. But it may well be possible for $; to be a proper subset of 52 and still have S; and Sp being of the same size in the sense that S) ~ Sp. For example, N-= {0,1,2,...,n,...} is equioumerous with the set E = {0,2,4,... ,2n,...)} of even integers, though E is a proper subset of N, Similarly, Q'~ A by the above, though Q is a proper subset ofuM Foundational problems A. Intuitively, any infinite set 5 is of the same size as some proper subset, while finite sets are just those which are not equinumerous with any proper subset So if S; is a subset of S; and we do not know whether 5, is a proper subset of Sz, one way to establish that is to show that Sy and Sp are not equinumerous, in symbols, 5; 7 Sai however, to do so would require ‘a special argument, since by the preceding S; ~ Sz is well possible for proper subsets. In the case of 5; = A and Sz = RR, that is accomplished by the results (3) and (4) above. Cantor's argument is navel not only in the concepts (set, equinumerosity, denumerability, nondenumerability) and results (1)-(4) involved but also in a basic point of methodology: Cantor finds a property P* of sets and shows two sets S;,5, to be distinct by showing that one of them has the property P* while the other does not Moreover, the existence of elements of a special kind follows by a purely logical argumnent: if S, is a subset of Sz and P*(S;) holds while P*(S2) does not hold, then S; must be a proper subset of Sa; that is, there exists an element x of $2 — $1 Since two finite sets {a1,a2,...,aq} and (b1,ba,-.. Pm} (of distinct elements) are equinumerous just in case they have the same number of elements (n = m), Cantor was led to say in general that two sets $; and So have the same numter of elements if $, ~ So. One may regard the number of elements in Sas an abstraction from its specific nature which isolates Just what it has in common with all equinumerous sets. ‘Thus, for example, {1,2,3} ~ (1,3,2} ~ {v2, ¥5, 7) all have the same number 3. For Cantor this was a process of double abstraction; the first level 5 abstracts away all that is distinctive about the elements of $ except how they are placed in a certain order, and the second level S abstracts away the order as well; 5 is called the cardinal number of S. Here instead we shall write card(S) for S$. For example card({ V2, /%,x}) = 3,card({5}) = 1, and curd(0) = 0. To identify the cardinal number of infinite sets we need new names. Cantor used the Hebrew letter 8 (aleph) with subscripts to name various infinite cardinal numbers, beginning with card(N) = Np. Thus also card(Z) = card(Q) = card(A) = No, but card(R) # Ny. To name the cardinal number of the continuum R, a new symbol ¢ is introduced by definition as ¢ = card(R). There is another way of naming card(R) that comes from the extension ‘of the arithmetic operations of addition, multiplication, and exponentiation to infinite cardinals, as follows. Suppose n= card(S,) and m = card $2); we can assume, without loss of generality, that Sy and S; are disjoint ‘Then define n+ = card(S; + Sz), where 51 + Sq is the union of S; and Sp. Next, define n x m = card(S; x S2) where S, x Sp is the set of all possible ways (x,y) of combining an element « of Si with an element y of 52. Finally, define n™ = card($$*) where $3* x SLX S consists of all possible combinations of elements of Si, one for each positionInfinity in mathematics in $2. For S1,S2 finite these definitions of +, x, and exponentiation agree with the usual ones, but for S; or Sp infinite we obtain new results. As ‘examples of calculations with the latter, one has m+ Ry = Ro-+No = No for any finite n. For this reflects the relations (01, mT (amt 1 dH 10,24). 9+ (13,57, 910,12... Thus unlike the case for finite m where n +m is greater than m (for n #0), we have m-+m = m for m = No. Similarly, for the same m we have m x m = m; this corresponds to the fact that a denumerable union of denumerable sets is denumerable. In order to represent the cardinal number ¢ of R. in terms of these operations, we return to the relationship of R with Q. Each element 1 of R is approximated by sequences or sets of rationals in various ways. One way is to associate with x the subset Qz of Q consisting of all rationals with r <2; then is uniquely determined by Qz. Let P(Q) be the set of all subsets of Q, in symbols P(Q) = {8 : S © Q}; thus cerd(R) < card(P(Q)).? Now for any set $ the set P(S) of all subsets 5” of $ is equinumerous with {0,1}, the set of all possible ways of associating a0 or 1 with each element + of $ (the correspondence puts 1 if the element ¢ belongs to S’ and 0 otherwise). Hence card(P(S)) = 2°°°4S) (and for this reason, P(S) is called the power set of $). In particular, card(P(Q)) = 2", +0, by the above, e< 2¥°. On the other hand, it is not difficult to produce 20 distinct members of R, by looking for example at real numbers 2 = O.ayazas,... in binary expansion, that is, where each a = 0 or 1. The conclusion is thus that (e= Obviously No < 2" since N is a subset of R; but since R is nondenu- merable, we must have (6) Ro < 2% ‘Thus there are at least two transfinite numbers No and ¢; can these be all? A Plethora of Transfinite Numbers Cantor was thus led directly (o the following questions: 1. Are there any infinite cardinal numbers besides Ny and e? Bor m= card{S),m = card(s) define n (QA -Q)] > P. Furthermore, the usual laws of quantification coupled with LEM yield (Bz) Pla) + >(¥2)-P(2) In other words, to establish an existence result (3r)P(z), it is suf ficient to assume (Vx)+P(x) and derive a contradiction. In general this will not show how to produce a solution « to satisly P(2). According to the platonist picture of set theory, then, statements such as the Continuum Hypothesis CH have a definite truth value, which it would be our aim as mathematicians to determine with all means at our disposal ‘The Foundational Issue (Part 2): Brouwer’s Way Off, Hilbert’s Way Out, and Weyl’s Way to Get By ‘The possible responses to those who found the platonist philosophy of mathematics unacceptable as a justification for higher set theory were ci- ther to reject that theory in whole or in part, or attempt some alternative form of justification. Here we review three basically different responses, due46 Foundational problems respectively to L. E, J. Brouwer, David Hilbert, and Hermann Weyl. To carry out their ideas, Brouwer and Hilbert created substantial foundational programs; these drew much attention but (as we shall see) proved far from successful. Weyl's initiative was more limited in its scope but achieved its aims within that; his basic work and approach is a forerunner of the results to be reported at the end of the essay [in chapter 12] Brouwer's Way Off Brouwer's program was the most radical and the most original of the three considered here. He advanced this with passionate convietion and persis- texice (despite its reception with general incomprehension and or rejection) from his doctoral dissertation in 1907 through his last paper in 1955.1° Brouwer strongly criticized platonism and formalism (of which, see below) In his 1908 paper “The unreliability of logical principles,” he argued that the Law of Excluded Middle for statements about infinite sets is obtained by an unjustified extension from finite sets, where it is correct. For exam- ple, when P is a testable property, we can verify (3z)P(z) V (Vz)-»P(2) for “2” ranging over a finite S, by inspecting each element of $ in turn, but this method is evidently inapplicable to infinite S. For Brouwer, questions of truth are restricted to statements that can be verified or disproved. Thus Py-P cannot be declared true until we have verified one or the other of its disjunets, Similarly (3x)P(x) cannot be recognized as true until we have found an instance a which makes P(a) true, On the other hand, to verify (¥z)P(z) in an infinite $ we cannot check each element of S, and other methods must be found according to the nature of §, Thus, for example, the Principle of Mathematical Induction, that P(O)A(Vz)[P(2) — P(+1}} implies (Vz)P(z), provides a method of drawing conclusions of the form (Ve)P(x) in the set N of natural numbers. Brouwer also rejected the com- pleted infinite, and he called collections $ such as Rand Cantor's second number class 1 (the countable ordinals} denumerably unfintshed: beyond any well-determined denumerable subset S" of such $ we can associate an clement of S— 5’. For Brouwer, then, the statement CH, which states the ‘equinumerosity of R and 9, has no definite meaning and the question of its truth has no interest for him. Brouwer thus took a completely constructivist stance in his critique of platonism, continuing the way advanced by Kronecker. But Brouwer gave his constructivism a particularly subjectivist stamp which he labeled in- tuitionism, emphasizing the origin of mathematical notions in the human intellect. Moreover, beginning in 1918, Brouwer went on to provide a sys- tematic constructive redevelopment of mathematics for the first time, going far beyond anything actually done by Kronecker Brouwer was Dutch and did hie dissertation in Amsterdam. All his published papers fon philosophy and the foundations of mathematics are to be found in volume 1 of his Collected Works (1975), the papers originally in Dutch are there translated into English. Incidentally, the second volume in his Cellected Works consists of papers in onconstructive mathematics, mainly topology, £0 which Brouwer made fundamental contributions during the period 1908-1913,Infinity in mathematics a7 ‘The theory of real numbers provided the first obstacle to a straightfor- ‘ward constructivist redevelopment. With real numbers identified as con- vergent sequences of rational numbers, all sorts of classical results would apparently fail to have reasonable constructive versions if one restricted attention only to sequences determined by effective laws. Brouwer intro- duced instead a novel conception, that of free choice sequences (Les.), which might be determined in nonlawlike ways (for example, by random throws of a die), and of which one would have only finite partial informa- tion at any stage. ‘Then with the real numbers viewed as convergent fc. of rationals, a function f from R to R can be determined using only a finite amount of such information at any given argument. Brouwer used this line of reasoning to conclude that any function from R to R must be continuous, in direct contradiction to the classical existence of discontin- ‘uous functions. With this step Brouwer struck off into increasingly alien territory, and he found few to follow him even among those sympathetic to the constructive position.” Hilbert’s Way Out When Hilbert addressed the International Congress of Mathematicians in 1900, he was nearing the age of forty and was already considered to be one of the world’s greatest mathematicians. Hilbert had by that time made his mark in algebra, number theory, geometry, and analysis. He would go on to make further substantial contributions in analysis, mathematical physics, and the foundations of mathematics. At Gottingen, where he was a profes- sor, Hilbert had many first-rate colleagues and students who often helped with the detailed development of his ideas. In particular, Hilbert’s program for the foundation of mathematics was taken up by von Neumann, Acker- mann, Bernays, Gentzen, and others, beginning in the 1920s. This program was shaped initially by both Hilbert's general tastes and interests as well as his specific experience with axiomatic geometry. Hilbert was noted for his clarity, rigor, and a penchant for systematically organizing subjects; at the same time he had the knack of putting these pedagogical tendencies to work to develop powerful methods for the solution of concrete problems. His work in algebra and algebraic number theory was part of the growing tendency toward abstract methods in the nineteenth century, which then came to dominate twentieth-century mathematics. One advantage of such methods is their generality—any mathematical structure meeting the basic principles must satisfy all the conclusions drawn from them. In his work on axiomatic geometry, Hilbert returned to the axiom sys- tem that had come down from Euclid, for which he gave a superior devel- opttient meeting modern standards of rigor. In addition, Hilbert moved on WWowadays, Brouwer's theories of Les. are much better understood and are demon: strably coherent, Dusnmett (1977) gives an excellent introduction tothe development of ‘mathematics based on Brouwer's intuitionistic ideas. (CT. also Troelstra and van Dalen (1988) and van Stigt (1990) |48 Foundational problems to what we would now call “metageometry,” through the study of questions, such as the independence of certain axioms from the remaining ones. This he achieved by a series of constructions of unusual models which satisfy all the axioms but the one to be shown independent. Also, Hilbert demon- strated the consistency of his axioms by the methods of analytic geometry, which interpret the statements in the Cartesian plane Rx R and space Rx RXR. In other words, geometry is shown consistent relative to a theory of real numbers (and for weaker combinations of axioms, relative to subsets of R such as the algebraic numbers). In the second problem of his 1900 addeess (Hilbert 1900, 1902), the compatibility of the arithmetical axioms, Hilbert proceeded to call for a proof of consistency for a system of axioms for R, which he recognized would, in some sense, have to be absolute. Hilbert’s statement of Problem 2 already reveals some of his key posi- tions, though they would be extended and elaborated later. He says there that the foundations of any science must be provided by setting up an ezact and complete system of axioms. “The axioms so set up are at the same time the definitions of [the] elementary ideas [of that science]; and no statement within the realm of the science whose foundation we are testing is held to be correct unless it can be derived from those axioms by means of a finite number of logical steps” (Hilbert (1902), in Browder (1976), p. 10}. After explaining the relative consistency proofs for geometry, he says: "On the other hand a direct method is needed for the proof of the compatibility of the arithmetical axioms” (ibid.) (that is, for a theory of real numbers). Hilbert goes on to posit that a mathematical concept exists if, and only if, it can be shown to be consistent (noncontradictory); thus, for him “the proof of the compatibility of the axioms [for real numbers} is at the same time the proof of the mathematical existence of the com- plete system of real numbers" (ibid., p. 11). The real numbers are not to be regarded as all possible convergent sequences of rational numbers, but rather as a structure determined by or governed by certain axioms, Fi- nally, in his statement of Problem 2, Hilbert expresses the conviction that the existence of Cantor's higher number classes will be demonstrated by a consistency proof “just as that of the continuum.” Evidently, at that time Hilbert thought the consistency proof of both the theory of real numbers and set theory would be straightforward. But within a few years (Hilbert 1904) he was presenting a less sanguine view: the paradoxes of set theory seemed to him to indicate that the problem of establishing the consistency of set theory presented greater difficulties than he had anticipated. Further- more, according to Paul Bernays, “although he strongly opposed Leopold Kronecker’s tendency to restrict mathematical methods, he nevertheless admitted that Kronecker’s criticism of the usual way of dealing with the infinite was partly justified.” TTBernays (1967), p. $00. For more information about Hilberts shifting views con- cerning foundations see Sieg (1984), pp. 166 and 170. [ef. also Hallett (1995).Infinity in mathematics 49 Tn the following decades Hilbert was preoceupied with the theory of in- tegral equations and mathematical physics, and he did not return to prob- lems of foundations until 1917." In the meantime, Zermelo had laid the axiomatic grounds for Cantorian (platonist) set theory while Brouwer had launched his attacks on mathematical platonism and formalism. Concern ing the latter, Brouwer stressed that consistency was not enough to justify the use of mathematical principles; what was necessary was to assure their correctness. Hilbert’s mature program for the foundations of mathematics via fint- {ist proof theory was announced in his 1917 address, “Axiomatic thought” (Hilbert 1918); it was then elaborated in a succession of almost yearly publications through 1931, Briefly, the idea is that a given body of nnath- ematies (such as number theory, analysis, or set theory) is to be treated as formally represented in an axiomatic theory T. Each such Tis to be specified by precisely described rules for generating its well-formed formu- las (statements) from a finite stock of basic symbols. Certain of these formulas are then selected as axioms (both logical ones and axioms con- cerning the specific subject matter of T), and rules for drawing inferences from the axioms ate specified. For a formal axiomatic theory T pre- sented in this way, the set of provable formulas is defined to consist of just those formulas for which there exists a proof (or derivation) in the system, that is, a finite sequence ending with the formula, cach term of which is an axiom or is obtained {rom preceding terms by one of the rules of inference. Then T is consistent just in case there is no contradiction (PAP) which is provable in T. Hilbert’s Beweisthearie, or theory of proofs, was developed as a tool to analyze all possible derivations in formal axiomatic systems. With proofs represented as finite sequences of formu- las, and formulas as finite sequences of basic symbols, whose structure in both cases is regulated by effective conditions, the question of consistency of T in no way assumes the actual infinite. Now Hilbert’s idea was to use entirely finitary methods in establishing the consistency of formal systems which otherwise required for their justification the assumption of the actual infinite. Not only the theory of real numbers but already a formal system of clementary number theory would have to be shown consistent. Such a sys- ‘tem would be a first-order version PA (Peano Arithmetic) of Peano’s axioms for N, formulated by replacing the second-order axioms of induction by the corresponding first-order scheme: P(0) A¥z(P(z) —+ P(x + 1)) > ¥rP(z), for all formulas P(x) in the language of PA. Hilbert wanted to justify the use of such a system including classical logic, which leads by LEM to statements like (3¢)P(z) V (¥z)~P(2); in this, he indirectly acknowledged Brouwer's criticism of the assumption of the actual infinite In addition to his general program, Hilbert proposed some specific proof-theoretie techniques to carry it out. These methods were shown [ft has been brought to may attention by W. Sieg that this is not quite co Hiller continued to lecture throughout that period on the foundations of mathemat ef Mallet (1995) ]50 Foundational problems to work in relatively simple cases for theories T much weaker than PA. ‘The success of the program would depend next on extending them to a, finitary proof of the consistency of PA, and so on, up to the consistency of set-theorctical systems like ZFC. As we shall see, these hopes were to be dashed by Gédel’s incompleteness theorems of 1931. Hilbert’s 1926 paper “On the infinite”'* is a very readable exposition of the finitist program for the foundations of mathematics and is a typ- ical example of Hilbert’s style of heroic optimism, which nowadays may be considered bombastic: “the definitive clarification of the nature of the infinite has become necessary, not merely for the special interests of the individual sciences, but rather for the honor of the human understanding itself” (Hilbert (1926), in van Heijenoort (1967), pp. 370-371). ‘The first portion of Hilbert's 1926 paper reviews the general “problem of the infinite” in mathematics, from which he turns to the particular prob- Jems raised by Cantor's theory of transfinite numbers. Here “the infinite was enthroned and enjoyed the period of its greatest triumph” (ibid, p. 375). But the paradoxes discovered by Russel! and others brought one to ‘an intolerable situation: is there no way to retain what Cantor achieved while escaping the paradoxes? Yes, by careful investigation and proof of the complete reliability everywhere of our inferences, “no one shall drive us from the paradise that Cantor created for us" (ibid, p. 375). The plan laid out in the mid-portion of Hilbert’s paper is that of ex- pressing mathematical propositions formally and representing mathemat- ical inferences by derivations in precisely described formal systems. The formulas of these systems are divided into finitary propositions and ideal propositions, Finitist proof theory is to be used to show how the latter can be eliminated in terms of the former; this is to be achieved by finitary proofs of consistency. The procedure of elimination here is analogous to that used to justify the introduction of idea! elements in mathematics such 4s imaginary numbers, points at infinity, and the algebraic number ideals introduced by Kummer. Apropos of this last, Hilbert remarks that “it is strange that the modes of inference that Kronecker attacked so passion- ately are the exact counterpart of what ... the same Kronecker admires so enthusiastically in Kummer’s work and praised as the highest mathemati- cal achievement” (ibid., p. 379). Moving on to more definite proposals, a formal system of elementary number theory is presented which is equiv- alent to the first-order axioms PA of Peano Arithmetic indicated above Hilbert claims that the problem of its consistency is “perfectly amenable to treatment”; moreover, what a “pleasant surprise that this gives us the solution also of a problem that became urgent long ago,” (ibid., p. 383) namely, the consistency of a system of axioms for the real numbers Finally, in the third part of his 1926 paper, Hilbert moves into high gear, leaving even his acolytes standing bewildered in the dust. For here 8We rler here tothe English translation in van Heijenoort (1967), pp. 369-392,Infinity in mathematics st he claims “to play a last trump” and to show how the continuum problem can be solved by use of proof theory, Hilbert's plan has intriguing as- pects roughly related to later work on constructive hierarchies of number- theoretic functions as well as to Gidel's own definitive results on the con- sistency of CH (to be discussed below), but it has never been worked out any coherent form." In his bravado and eagerness to demonstrate the success of his program, Hilbert. promised far more than he could deliver. Indeed, even the claims of finitist proofs of the consistency of elementary rnumber theory and real number theory would not survive Gédel's incom- pleteness theorems of 1931. For one who trumpeted the cause of absolute reliability, how wrong could he be? Nevertheless, Hilbert's influence and program were decisive factors in what followed: a5 we shall see, the attack fon the “problem of the infinite” was shifted to the arena of metamathe- matics and in that proof theory became one of the primary tools. One final remark concerning Hilbert vis & vis the constructivists needs to be made. As Paul Bernays tells it, “there was a fundamental opposition in Hilbert’s feelings about mathematics . .. namely his resistance to Kro- necker’s tendency to restrict mathematical methods . . . particularly, set theory . . . and his} thought that Kronecker had probably been right. It became his goal to do battle with Kronecker with his own weapon of finiteness." But in his efforts to outfight the constructivists, Hilbert. was hoist with his own petard Weyl’s Way to Get By Hermann Weyl was one of Hilbert's most illustrious students, and his work was almost as broad (and deep) as that of his teacher. After a period as Dozent in Gottingen following his doctoral work there, Weyl took & position in Ziirich shortly before World War I. Hilbert made several efforts to bring, him back, and Weyl finally returned as Hilbert’s successor in 1930, only to leave for Princeton when the Nazis came into power in 1933. All these personal and professional connections made it difficult for Weyl to express his strong differences with Hilbert on foundational matters, but he did so tactfully on a number of occasions.2” ‘Weyl evolved his first approach to the foundations of mathematics in the monograph Das Kontinuum (1918). In the introduction thereto he criticized axiomatic set theory as a “house built on sand” (though the ob- jects of, and reasons for, his criticism are not made explicit). He proposed to replace this with a solid foundation, but not for all that had come to [GF the discussion by R. MI. Soloway of ene of Gadels proofs of the consistency of (CH in Gadel (1998), 114-127] Quoted in Reid (1970), p. 173. Weyl’ best known works on logic and the philosophy of his mathematics are his monograph (1918) and book (1949), both ditcussed Below, But there are also many lest familiar article on these subjects seatered through his Collecel Works (1968), spat ting the period 1918-1955. [Chapter 13 in this volume provides 8 detailed examination Of Weyl (1918); age also chapter 14]52 Foundational problems be accepted from set theory; the rest he gave up willingly, not seeing any other alternative. Weyl's main aim in this work was to secure mathematical analysis through a theory of the real number system (the continuum) that ‘would make no basic assumptions beyond that of the structure of natu- ral numbers N. Unlike Hilbert, Weyl did not attempt to reduce first-order reasoning about N to something supposedly more basic. In this respect Weyl agreed with Henri Poincaré that the natural number system and the associated principle of induction constitute an irreducible minimum of theoretical mathematics, and any effort to “justify” that would implicitly involve its assumption elsewhere (for example, in the metatheory). And, unlike Brouwer, Weyl accepted uncritically the use of classical logic at this stage (though at a later date he was to champion Brouwer's views). For- ‘mulated in modern terms, a system like PA was thus accepted by Weyl as basic. However, for a theory of real numbers one would have to provide a ‘means to treat sets or sequences of natural numbers (using the reduction of Q to N) and then, for analysis, explain how to deal with functions of real ‘numbers as functions from and to such sets, Weyl added axioms for exis- tence of sets of natural numbers which are arithmetically definable; these are of the form (3X ¢ N)(¥n € N)[n € X = P(n)] where the formula P(n) contains n0 quantified variables other than those which range over N. Using two sorts of variables, lowercase for elements of N and uppercase for subsets of N, these axioms take the form (3X)(Vn)[n € X = P(x). Weyl deliberately excluded axioms of this type in which P contains quan- tified variables ranging over subsets of N; in particular, he thus excluded statements of the form (3X)(¥n}[n € X + (WY}Q(n.¥)], even when Q is an arithmetical formula. Weyl's reason for doing so was that otherwise one would be caught in a circulus vitiosus. The matter at issue here requires a lengthy aside, ‘The vicious circle principle, first enunciated by Poincaré, was designed to block certain purported definitions, in which the object introduced is, somehow defined in terms of itself. According to Poincaré, all mathemat- ical objects beyond the natural numbers are to be introduced by explicit, definitions. But a definition which refers to a presumed totality—of which the object being defined is itself to be a member—involves one in an ap- parent circle, since the object is then itself ultimately @ constituent of it ‘own definition. Such “definitions” are called impredicative, while proper definitions are called predicative; put in more positive terms, in the latter one refers only to totalities which are established prior to the object being defined. The formal definition of a set X of natural numbers as X = {n (¥Y)Q(n,¥)}, taken from the axiom 3Xnjn €X 4 (VY)Q(m,Y is im- predicative in Poincaré’s sense because it involves the quantified variable “¥” ranging over arbitrary subsets of N, of which the object X being de- fined is one member. ‘Thus in determining whether (¥¥)Q(n, ¥) holds, we shall have to know in particular whether Q(n, X) holds —but that can’t be settled until X'itself is determined.Infinity in mathematics 53 Poincaré raised his ban on impredicative definition thinking doing so would exclude the paradoxes. However, that succeeds only by taking a very broad reading of what it means for a definition to refer to a totality. For example, in Russell's paradox, derived from Javz(z € a ++ x ¢ 2}, or a= {r:2¢ 2}, the formula P(x) = (e ¢ 2) does not refer to the pre- ‘sumed totality of all sets, since it does not contain any quantified variables ranging over sets. On the other hand, it is true that in order to deter- mine by its “definition” which members a has, we must already know the answer; in particular, a € a would be determined true only if we already knew that a ¢ a. However, this vicious circle is not itself excluded by the principle enunciated by Poincaré in its prima facie reading. On the other hand, Poincaré’s principle as it stands would certainly exclude impredica- tive definitions in analysis of the sort X = {n : P(n)} where P contains quantified variables ranging over arbitrary subsets of N; the objection to such definitions is not that they are paradoxical, but rather that they are implicitly circular and hence not proper. Russell was one of the first to accept Poincaré’s ban on impredicative definitions as applied to sets, and he turned to the construction of a for- talism for generating predicative definitions. Roughly speaking, instead of talking about arbitrary sets, one may only talk in this formalism about sets of level 1, sets of level 2, etc. A set of level 1 is defined by a formula using quantified variables ranging over individuals only. ‘This determines the totality of sets of level 1; then a set of level 2 is defined using just quan- tified variables ranging aver individuals and over arbitrary sets of level 1 ‘The problems with this kind of ramification (as Russell called it) is that in the resulting theory of real numbers (introduced as sets of rational num bers) one would have reals of level 1, reals of level 2, ete., but analysis with such distinctions would be unworkable. As an ad hoc device, Russell then Introduced his Axiom of Reduciility, which says that every set of higher level is coextensive with one of lowest level. In effect, this “axiom” nullified the point of ramification and indirectly permitted the use of impredicative definitions in analysis. Russell recognized that this step did not jibe with the initial philosophical outlook stemming from Poincaré’s definitionism, but saw no alternative to developing analysis. As it happens, his theory of types, which distinguished sets according as to whether they were sets of individuals (type 1), sets of such sets (type 2), etc., and which restricted the membership relation to objects of successive types (with individuals taken to be of type 0), already blocked the paradoxes without in forcing one to insist on predicative definitions. This was achi later move (due to F. P, Ramsey) to Simple Type Theory as opposed to Russell's original Ramified Type Theory (which distinguished sets both as to types, and as to levels of definition). Now if one accepts the platonist philosophy in set theory, the totality P(N) of subsets of N exists independently of how its objects may be de- fined, if all. According to this view any formula P(2) involving variables ranging over N and variables ranging over P(N), connected by arithmeticalcy Foundational problems relations between individuals and relations of membership between individ- uals and sets, has a definite meaning, independent of whether we have any means to “determine” it; then the definition X = {x : P(z)) simply serves to single out one member of P(N), According to the platonists, this is entirely analogous to singling out @ natural number as the minimum one satisfying a certain (nonempty) property, formally k = (min n)P(n), where wwe may have no way to determine k effectively, even when P only involves numerical quantification. But according to the predicativist (or “defini- tionist”) there is a difference: the totality of natural numbers is granted as clear and definite and each of its members can be singled out by a prior representation, while there is no justification in the assumption of a totality of subsets of N independent of how these may be defined, since sets can be introduced only by definition on the basis of (successively) established totalities ‘To return to Weyl, we can say that he positioned himself as follows in Das Kontinuum2 First, he rejected the platonist philosophy of mathe- ‘matics as manifested in the impredicative existence principles of axiomatic set theory, though he accepted classical quantifier logic when applied to any established system of objects. Second, he accepted the predicativist viewpoint taking the system of natural numbers as its point of depar- ture. Third, he recognized that, whatever their justification on predica- tive grounds, ramified theories would not give a viable account of anal- ysis. Weyl’s main step, then, was to see what could be accomplished in analysis if one worked just with sets of level 1, in other words, only with the principle of arithmetical definition. Here it is to be understood that such definitions may be relative; that is, if P(z,¥i,...Yq) contains vari ables ¥i,... ,Y_ but no quantified set variables, then P serves to define X = {x! P(2,Yi,... .Yq)} relative to ¥i,...Y¥q- Given any specific defini- tions of ¥i,..- ,Ya this will produce by substitution in P a specific defini tion of X. Finally, such means of relative arithmetical definition serve to explain which functions from sets to sets are to be admitted, namely, just those given as F(Y;,... »Yq) = {e+ P(x,Yi,-.. »Ya)} with P arithmetical In this way, Weyl was able to set up a formal axiomatic framework for arithmetical analysis in which he could define rational numbers in terms of (pairs of) natural numbers, then real numbers as certain sets (namely, lower Dedekind sections) of rational numbers, and finally fanctions of real numbers as functions from sets to sets in the way just explained. He then went on to examine which parts of classical analysis could be jus- tified on these grounds. To begin with, the general least upper bound (Iu.b,) principle for sets of real numbers could not be accepted in its full generality. According to that principle, any set Sof real numbers which is bounded above has a Lu.b. X. In Weyl’s framework, 5 is given as consisting of lower Dedekind sections ¥ satisfying an arithmetical con- Tigee also his elaboration ofthis position in Wey! (1919).Infinity in mathematics 55 dition P,(Y); then its supposed 1.u.b. X is the union UY{P)(Y)); that is, X = (r € Q: 3V(R(Y) Ar € ¥)}._ Translating rational numbers to natural numbers, this gives X = {n : (3Y)P(n, Y)} for suitable arith- metical P; but this is an impredicative definition which is not justified in Woyl’s system. On the other hand, the Lu.b. principle for sequences of real numbers does hold in his system. For, a sequence (Yé)en is given by an arithmetical condition F(r,&), which holds just in case r € Ye. If the sequence is bounded above, its Lu.b. X is the union UY,[& € N), that is, X = {r €Q : (3k)P;(r,k)}, and this reduces to an arithmetical definition of the form X = {n: (3k}P(n,k)} with P arithmetical?® Thus Weyl’s task was reduced to seeing which parts of classical analysis rest simply on the Lu.b. principle for sequences of reals rather on the more general 1u.b. principle for sets of reals. Here, in fact, he found that the entire theory of continuous functions of reals goes through in a straightforward manner in arithmetical analysis, including for such (i) the intermediate value the- orem, (i) the attainment of maxima and minima on a closed interval, and (iii) uniform continuity on a closed interval? From these it is a direct step to develop the theory of differentiation and integration for continuous functions. Here it is Riemann integration that can be treated in a straight- forward predicative way. (Weyl remarks® that it is less simple to deal with the more modern theories of integration through theories of measure, but gives no indications; their treatment would be left for the future develop- mts of predicative analysis outlined in chapters 12-14 in this volume].) There is no mention of Brouwer or intuitionism, and no restriction on the logic employed, in Das Kontinuum, Howover, within the next few ‘years Weyl became more familiar with Brouwer's work and somewhat of a convert. He is quoted as saying, during some lectures on Brouwer's ideas in 1920: “I now give up my own attempt and join Brouwer." Over the following years Weyl was to help champion Brouwerian intuitionism as against Hilbert's program in various publications, much to Hilbert’s annoy- ance. However, in later years he became pessimistic about the prospects for the Brouwerian revolution. Moreover, it is not true to say that Weyl ever completely gave up his “own attempt,” which he continued to men- tion over the years in articles on foundational matters, where his criticisms of mathematical platonism in set theory remained constant. A relatively mature expression of Weyl's view is provided by his 1949 book, which is Weyl establishes Cauchy's convergence principle for sequences of real, which is equivalent to the Lu.b. principle for sequences. Note iso that the | b.priniple for sets (sequences) i equivalent to the g1b. principle forthe same, and with the given recentation of real mubers this leads to definition of the gb. as an intersection rather than a union, 28s sketched in Wey! (1918), pp. 61-65, [Cf also chapters 13 and 14 in this volume] ibid p65 28See van Heljenoort (1967), p. 480; the original statement may be found in Weyl Gesarmmelte Abhandlungen (1968), vol. 2, p. 15856 Foundational problems readily accessible.?® The following quotations from this book reveal both. his settled convictions as well as his unsettled state of mind about the eventual foundations of mathematics: ‘The leap into the beyond occurs when the sequence of numbers that is never complete but remains open toward the infinite is ‘made into a closed aggregate of objects existing in themselves. Giving the numbers the status of ideal objects becomes dan- gerous only when this is done ‘The vindication of this transcendental point of view forms the central issue of the vio- lent dispute . .. over the foundations of mathematies.2” ‘From an aggregate of individually exhibited objects we may by selection produce all possible subsets and thus make a survey of them one after another, But when one deals with an infinite set like N, then the existential absolutism for the subsets becomes still more objectionable than for the elements.#* ‘There follows an explanation of the vicious circle principle, of levels of pred- icative definition, and of the resulting “dilemma” for analysis concerning the Lu.b. principle: Russell, in order to extricate himself from the affair, causes rea~ son to commit harakiri, by postulating the above assertion [the Axiom of Reducibility] in spite of its lack of support by any ev- idence. . .. In a little book Das Kontinuum, published in 1918, Thave tried to draw the honest consequence and constructed a field of real numbers of the first level, within which the most important operations of analysis can be carried ont.° Mathematics with Brouwer gains its highest intuitive clarity. He succeeds in developing the beginnings of analysis in a nat- ural manner, all the time preserving the contact with intuition much more closely than had been done before. It cannot be denied, however, that in advancing to higher and more gen- eral theories the inapplicability of the simple laws of classical logic eventually results in an almost unbearable awkwardness. And the mathematician watches with pain the larger part of his towering edifice which he believed to be built of concrete blocks dissolve into mist before his eyes. Finally, “Hilbert’s mathematics may be a pretty game with formulas . . but what bearing does it have on cognition, since its formulas admittedly What is a revised and considerably augmented English edition of an article Weyl wrote forthe Handbuch der Philosophie in 1926 2 Weyl (1948), p38. "bid, p49. 27d, p80. Smid, p54Infinity in mathematics 57 have no material meaning by virtue of which they could express intuitive truths?" In this connection, Weyl says that a consistency proof of arith- metic “would vindicate the standpoint taken by the author in Das Kontin uum, that one may safely treat the sequence of natural numbers as a closed sequence of objects.”%? Incidently, a famous wager was made in Zirich in 1918 between Weyl and George Pélya, concerning the future status of the following two propo- sitions: (a) Each bounded set of real numbers has a precise upper bound. (2) Each infinite set of real numbers has a countable subset. ‘Weyl predicted that within twenty years either Pélya himself or a ma- jority of leading mathematicians would admit that the concepts of number, ‘set, and countability involved in (1) and (2) are completely vague, and that it is no use to ask whether these propositions are true or false, though any reasonably clear interpretation would make them false (unless the concepts involved were to acquire totally new meanings). The loser was to publish the conditions of the bet and the fact that he lost in the Jahresberichten der Deutschen Mathematiker Vereinigung, this never took place as such.* Weyl’s viewpoint in making this wager is often mistakenly identified as being that of Brouwer’s intuitionism, though it was made at the time of publication of Das Kontinuum, prior to Weyl's taking up Brouwer's views. ‘As we have seen, the L.u.b. principle was rejected in his 1918 publication on the grounds of the vicious circle principle and the rejection of impredicative definitions. And (2) requires some form of the Axiom of Choice applied to subsets of R for its proof, so was rejected by Weyl along with his rejection of the conception of R as a completed totality. As to the settlement of ‘the wager, there is no question that Weyl lost under its stated conditions. Moreover, the vast majority of mathematicians (leading or otherwise) would say then, as they would say now, that Weyl’s side of the bet was simply wrong-headed. But Weyl would not be shaken by this: “The motives are clear, but belief in this transcendental world taxes the strength of our faith hardly less than the doctrines of the early Fathers of the Church or of the scholastic philosophers of the Middle Ages." So, the “Middle Ages” are Wey) (1939), p 61 31/6id, p60," Por elaboration of Weyl's views, the reader should not overlook Ap: pendix A of Weyl (1949), and see further his 1944 obituary article “David Hilbert and his mathematical work,” reproduced in Red (1970), pp. 245-285, ax well as Wey! (1948) ICE alo chapter 13 in thie volume | "he document spelling out the bet was reproduced by George Pélye in » brief note (1972). According to Pélya, "The outcome of the bet became a subject of discussion between Weyl and mea few years after the fnal date, around the end of 1940. Weyl thought that he was 49% right and I, 5196; but he also asked me to venive the conse quences specified in the bet, and I gladly agreed." Palya shoved the wager to many fiends and colleagues and, with one exception, all thought that he had won Weyl (1946), p. 658 Foundational problems simply taking somewhat longer to wane than Weyl expected—but that's the way it goes with Middle Ages. The Rise of Metamathematics and the Crumbling of Hilbert’s Program Hilbert’s emphasis on the axiomatic approach to mathematics and on the metatheoretical questions concerning axiom systems constituted one of the principal sourees for the development of metamathematics as a new, istinctive, and coherent subject. Here “meta” has taken the meaning “about”; that is, axiom systems are objects of study examined externally. Such a treatment fits very well with axiomatic mathematics in its tradi- tional sense, as first exemplified in geometry, but also with modern axioma- tizations of number theory, algebra, and topology; each of these deals with a restricted or “local” part of mathematies. The metamathematical ques- tions that Hilbert raised about axiom systems concerned their consistency, completeness, categoricity, and independence (of the axioms, one from at other). In his program for the foundations of mathematics via finitist proof ‘theory, Hilbert imposed further requirements on how the investigation of such questions was to be carried out, but those restrictions are not essential to metamathematics and only influenced part of its development. What ‘s essential is that axiom systems are to be described precisely in formal terms so that results concerning them may be established rigorously. The term “metamathematics” then is suitable for the study in each case of that part of mathematics formalized in a given axiom system. But this is already troublesome, insofar as such study is to be carried out, by infor- mal mathematical means which may themselves be formalized. This novel feature is exactly what was capitalized on in the first striking results of metamathematies, namely, the Lowenheim-Skolem theorem and Gédel's incompleteness theorems; these, furthermore, combined to undermine both Hilbert's conception of mathematical existence and his finitist program for establishing “existence” via consistency proofs ‘The following is devoted primarily to the highly discomfiting (if not devastating) effects on Hilbert’s program of several metamathematical re- sults of a general character (including those just mentioned). At the same time, these results served to undermine the programs for the universal or “global” axiomatization of mathematics initiated by Frege and carried on in the work of Russell, Zermelo, Fraenkel, and others. In contrast to the ax- iom systems for restricted areas of mathematics of the sort described above, here one aimed at systems of such generality that “all” mathematics could be formalized within them. As such, the idea of investigating these systems from the outside was antithetical to the motives for their creation,?® but as 35 This point has been emphasized particularly by van Heijenoort; ef, for example, his 1967 essay reproduced in van Hejenoart (1980), pp. 13-14Infinity in mathematics wwe shall see, the claims to universality could not withstand the onslaught of metamathematics. ‘The Léwenheim-Skolem Theorem In 1915 Leopold Léwenheim established a theorem which is now stated as follows: if a formula A of the first-order predicate calculus has a model (is satisfiable in some domain), then it has a denumerable model. ‘This ‘was improved by Thoralf Skolem in 1920 (and again in 1922) both in the proofs and in the statement of results. Skolem’s theorem tells us that if S is any set of first-order formulas which has a model, it has a denumerable model. The Léwenheim-Skolem theorem was derived later (1930) a5 a consequence of Gédel’s completeness theorem, by which if $ is consistent im first-order predicate logic, then $ has a denumerable model, For it is trivial that if $ has a model at all, then it must be consistent.2° At first sight, Gadel’s completeness theorem would seem to support Hitbert’s dictum that existence of a mathematical concept is the same as consistency of an axiom system for it. But Hilbert had in mind axiomatiz tions S like that of Peano for the natural numbers and of Dedekind for the reals (a8 well as certain of his own for geometry) which are categorical, that js, such that the models of the axiom system are uniquely determined up to isomorphism. This implies that if M,/M’ are any two models, then they are equinumerous; that is, M ~ M’. But the Lowenheim-Skoiem theorem shows that no set of first-order axioms $ for the real numbers, that is, for which R is a model, can be categorical since $ has a denumerable model while R is nondenumerable. A later result by Skolem showed that, no set of first-order axioms for IN can be categorical; he did this by constructing a nonstandard model 1’ satisfying exactly the same first-order statements as N.2" A still more general theorem of Tarski showed that if a first-order set of statements $ has a model M of any infinite cardinality m, then it has a model M’ of any other infinite cardinality m’, So no set of first-order axioms $ with an infinite model can be categorical Since Peano’s original axioms for N and Dedekind's for Rare categori- cal, they must have an essentially non-first-order component. In fact these are just, respectively, the axioms of induction for N and of completeness Whe various Basie papers of Lowenheim, Skolom, and Gidel may be found translated in van Heijenoort (1967), [The paper of Sholom dated 1922 in that source was not published until 1923 and appears a8 Skolemn (1924) in ou bibliography: the reason given by van Heljenoort forthe earlier dating is that Skolem had lectured on this materia in 1922}} See also Cadel (1986). [Chapter 6 in this volume provides e survey of Gisdel's ie and work "rin the language of caren mode hen, to structures AM" hich provi in terpretations of the same language are called elementary equivalent, and one Writes ALE M’, stthey satisfy the same fes-order statements in that language. Thea hy = nonstandard model of the natural numbers is meant an N’ not isomorphic to N with NSN". Similarly a nonstandatd model ofthe resis an R such that R= A! but Rt is not isomorphic to R60 Foundational problems (or the Lu.b. principle) for R. expressed set-theoretically. For example, in the case of Peano’s axioms this takes the form VXI0 EX AVn(n € X + (n+ 1) € X) + Yn(ne XI}, where the (set) variable “X” ranges over subsets of the domain. But for categoricity there is a further requirement, called that of standard second- order logie: in any interpretation of the axioms in a domain M, the set variables are to be interpreted as ranging over the set P(M) of arbitrary subsets of M. Now this takes on a more puzzling aspect when we move to axioms for sets like those of Zermelo-Fraenkel, ZF (or even Zermelo’s system Z) Since ZF is first-order, if it has a model M at all then it has a denumerable model M'. But any set 4 in M’ must have an interpretation ofits powerset P(A) in M’, say Pyy(A). In ZF it is a theorem that if A is infinite then P(A) is nondenumerable. But in Mf’, Pyy(A) has only denumerably many members. This peculiar situation is referred to as “Skolem’s paradox” though it is not actually paradoxical; the puzzle is resolved by noting that externally, A ~ Pyy-(A), since both sets are denumerably infinite. But internally, that is, within M’, there is no function which can establish the one-to-one correspondence between A and the interpretation of P(A) in -M’. In other words, the formal statement =(4 ~ P(A)) is still true in M’ Now if we are to follow Hilbert’s dictum of mathematical existence through in this context as well, the existence of set-theoretical concepts rust be secured by the consistency of axiom systems like ZF. As we see, however, what would thereby be secured is not the intended concept but 2 variety of possible interpretations in models of different cardinality. The only way to secure the concept of set axiomatically is to add to the axioms of ZF a second-order “completeness” axiom referring to arbitrary subsets of the intended model M. ‘This now puts us in a bind: one can’t capture the notion of arbitrary set axiomatically without somehow already assuming ‘the notion of arbitrary set to be understood ‘There is still a final problem for Hilbert’s conception of mathematical existence when combined with his finitist program. According to the latter, each axiomatic system must be prescribed finitarily: the syntax, axioms, and rules of inference are all to be determined by finitary effective proce- dures. For example, this is achieved in the axiom system PA by replacing Peano’s second-order axiom of induction by the axiom scheme consisting of all formulas of the form P(O)AVn(P(n) + P(n + 1)) + ¥nP(n), where P(n) is any formula in the first-order language of arithmetic using, the basic symbols (0,1,+,-, and =). Now from a purely formal point of view, the statement of Peano’s original axiom, YXIOE X Ann € X + (n+1) EX) Vain eX),Infinity in mathematics a appears to be equally finitary. Indeed this is so, but being so does not suffice for categoricity; there is nothing in such a formal axiomatization within a second-order language that can force the system to be categorical Only the semantic requirement that we are dealing with standard second- order logic, that is, that in any interpretation M, “X" ranges over P(M), will ensure categoricty. But that requirement is on its face nonfinitary; in fact, this is demonstrated as a consequence of the results with which we deal next. Gidel’s Incompleteness Theorems In his famous paper “On formally undecidable propositions of Principia ‘Mathematica and related systems,” Gédel (1931) proved that for a wide class of effectively presented consistent. extensions S of a. formal system PM containing PA, there are elementary statements A such that neither A nor “A is provable in S, For his proof, Gédel introduced a class of finitarily defined functions which are obtained from the zero and successor functions by explicit definition and inductive definition on N.*® The class of functions, so generated is now called the primitive recursive functions. All such are effectively computable; through later work of Church, Kleene, Turing, and Post,2® one arrived at a definition of the most general class of effectively computable functions on N, which is now called the recursive junctions, Gdel’s incompleteness results were achieved by first coding up syn- tactic objects (expressions, terms, formulas, proofs) as sequences (or se- ‘quences of sequences, ete.) of numbers and then representing such sequences (m,... ,me) as single numbers n using the prime power representation Dit. pk (this process is called Gédel numbering). Gédel then used this numbering to formally represent the syntax of the system PM within itself, More generally, he showed that if S is any extension of PM with & primitive recursive set of axioms (under its Godel numbering) and S is, (what he called) w-consistent, then S is incomplete; that is, there are state- tents A such that neither A nor “A is provable in S. Here w-consistency is a mild technical extension of the usual concept of consistency. ‘The hy- pothesis of «-consistency was later replaced by that of simple consistency by an argument of Rosser,"° who further strengthened Gddel’s theorem so as to apply to any recursive system S extending PM. Gédel pointed out in 1931 that much weaker systems than PM served the same purpose, In fact, other later improvements gave the result that any recursive consistent extension S of PA is incomplete. Win the simplest case, of a function of one arguinent, inductive definition sakes the form F(0) = ¢.F{nt 1) = H(n,P(n)), where HT is a previeusly defined func- tion; more generally, one defines functions F(n,ms,... sma) by P(O,m3,... me Glenys), FO Ay mya) = Hmm, M4, FM... sm) where G and Hi are previously defined functions. Nowadays, one uses “recursive” for "inductive" in auch definitions of functions "See Davis (1965) $8See his 1937 paper reproduced in Davis (1966) “\See Tarski, Mostowski, and Robinson (1953) for the history and still more general62 Foundational problems One can say more about the kinds of statements A which are not decided by such S; they can be chosen in the form ¥n(F(n) = 0) where F is a primitive recursive function. For primitive recursive S, Gédel arrives at this by showing that the relation Proofs(m,n) which expresses that n is the number of & proof in $ of the statement with number m, is primitive recursive; that is, it has a primitive recursive characteristic function. He then constructs by a diagonal procedure a statement A such that a) A ¥n-Proofs(!Al,n) is provable in S, where ‘4? is the number of 4; informally speaking, A expresses of itself that it is not provable in $. By the above we obtain a primitive recursive function F such that F(n) = 0 ++ +Proofs(‘A',n), so 2) Ao ¥n(F(n) = 0) For this choice of A, Gédel shows that (3) (i) if Sis simply consistent then A is not provable in S, and (ii) if $ is w-consistent then +A is not provable in S, This is Godel’s first incompleteness theorem. (The construction of A must be modified for Rosser's improvement to simple consistency in (i) The consistency of S can be expressed as the statement that no proof in Sends in a contradiction; this may also be expressed by a statement Cons of the form ¥n(G(n) = 0), with G primitive recursive. By ordinary logic, consistency is equivalent to the statement that some formula is not provable in S; hence the statement YnProofs('A',n) implics Cons. Now {8)(i) expresses informally that the following formal statement is true: (a) Cons = Yn-Proofs('A",n). Gédel succeeded in showing that not only is (4) true, but itis also provable in S—this, by formalizing the argument for (3)(i). Combining all of these facts it follows that (5) if Sis simply consistent then Cong is not provable in $ (Gédet’s second incompleteness theorem). “These results were both stunning and disappointing. Consider the latter aspect first: since, under the given hypothesis, the A constructed in (1) is not provable in S, then A is obviously true, for that is just what A expresses of itself. No statement of prior mathematical interest (such as the twin prime conjecture or the Riemann Hypothesis) has been shown results. The reduction to PA essentially makes use of Gédel's result from his 1931 paper that all primitive recursive functions are arithmetically definable, that is, definable in the first-order language having 0,1, +,-,= a8 basic symbolsInfinity in mathematics 63 to be independent of S. Rather, the statement produced has simply been cooked up for the occasion, again by a diagonal trick. Even more, since A is obviously true, there is no question which of A, A we would add to S as an axiom if we were to try to overcome this instance of incompleteness; if is to reflect truth in N then A should be added. Of course then S' = S +A} is again subject to the incompleteness theorem, but that is another matter. What was stunning about the incompleteness theorems was how they put Hilbert’s entire conception of mathematical existence and his consis- tency program into question. First of al, if one gives up categoricity as the requirement on an axiom system S for it to uniquely specify the concept of ‘a mathematical structure (such as that of the natural numbers), still one might hope that all truths in the supposed structure would be derivable in S, in other words, that $ would be complete, Indeed, Hilbert had said as much in the statement of Problem 2 in his (1900) list of problems (quoted above), and he later specifically conjectured that a formal system for arith- metic is complete. This was shown false by Gédel’s first incompleteness theorem, Second, Hilbert had divided the propositions of a formal system into those which are real and those which are ideal, the latter being used simply fas an aid to derive the former. Among the “real” propositions would be all statements of the form Yn(F(n) = 0) where F is a finitarily defined function. Since Hilbert granted that all primitive recursive functions are finitarily defined, Gédel's results from (2) and (8) above showed that no one consistent formal system S could even serve to derive all true “real” propositions. Thus, even the fallback position fom completeness for all statements to completeness only for the class of “real” statements gives way. Third, Hilbert’s call for proofs of consistency of axioms systems to se- cure the concepts they were supposed to express fails to secure the correct ness of some of the most elementary theorems in those systems. For if A is a true statement of the form ¥n(¥(n} = 0) with F primitive recursive which is not provable in S, then $ +{4} is consistent and proves the false statement 3n(F(n) #0). Still it is at least the case that consistency of S (of the sort to which the incompleteness theorems apply) ensures that if S proves a “real” statement B of the form ¥n(G(n) = 0), with G primitive recursive, then B must be true; for, otherwise, there would exist a specific k such that G(k) # 0, hence also 3n(G(n) # 0) would be provable in contradicting ¥n(G(n) = 0). Finally, there are the unsettling consequences that Godel’s second in- completeness theorems ((5) above) have for Hilbert’s program of finitary Whi was pointed ow by Godel in hie fist public announcement (in 1930) of the Sncompleteness Uneorems; soe Godel (1986), pp. 196-204 ‘©The point has been stressed particularly by G. Kreisel. See his article (1978) on Hilbert's Second Problem, in Browder (1976)4 Foundational problems consistency proofs. Since all mathematics is supposed to be formalizable, this holds in particular for finitary mathematics, Now if all fnitary methods can be represented in a single formal system S, then either $ is not consis- tent or its consistency cannot be proved by finitary methods. A possible alternative is that no one consistent formal system can serve to formalize all of finitary mathematics, though any given portion can be formalized in ‘one system or another. However, itis dificult to conceive how single global systems like PM or ZF could fail to formalize all finitary arguments, In fact, all the finitary proofs of consistoncy of various systems that had been cartied out by workers in the Hilbert school were already formalizable in PA, and this explains why their best efforts failed with PA itself. Never- theless, it remained conceivable that the use of finitary methods involving novel and/or more difficult. arguments could serve to handle PA and even- tually go on to succeed with systems like PM and ZF. Gédel himself was cautious on this point at the conclusion of his 1931 paper, but eventually he came to the view that Hilbert’s conception of fnitism is limited by PA.** A basic problem here is that the general concept of finitary proof is not sufficiently clear that one can draw definitive conclusions about its possible limitations. Every actual finitary proof which had been carried out could be recognized as such, and most proofs of current mathematics are evi- dently nonfinitary, but this experience has not been sufficient to determine the concept in any sharp way. Beginning with Gerhard Gentzen in 1936,4° workers in proof theory extended the evidently finitary methods employed for various consistency proofs to include certain éransfinite elements, such as principles of induction on various effectively presented systems of ordinal notations, while hewing to Hilbert’s requirement to use only “real” state- ments ¥n(F(n) = 0), with F primitive recursive. (Cf. chapter 10 below.} Postscript on Second-Order Logic ‘Gédel’s 1930 completeness theorem showed that an effectively given system PC; of axioms and rules of inference for the first-order predicate calculus serves to prove just those statements which are valid in all models. His 1931 results could be used to show that there is no corresponding completeness theorem for second-order predicate calculus PC2, if by validity is meant validity in the standard sense, that is, that in each interpretation M the set variables required are to range over P(M), Suppose to the contrary that PG» has an effectively given system of axioms and rules which proves Just those statements that are valid in this sense. Then Peano’s axioms, with induction in its set-theoretical form are categorical when we use with them the logic of PCa; call this system PA2. Now since PA is true in the structure of natural numbers, every theorem A of PAs is a truth about IN. Conversely, every truth A about N is equally a truth about any model ‘See the introduction to Godel (1958); this paper is reproduced in translation, to- gether with a later extended, revised version (1972) in Gisdel (1990). ‘8See Gentzan (1969),Infinity in mathematics 65 isomorphic to N, hence any model of PA; so, finally, A is a theorem of PAg, ‘Thus PAg is formally complete: for every statement A in its language, either PAz proves A or it proves +A. Since PAz contains PA, and is supposed to be effectively given, the conclusion violates Gédel’s first incompleteness theorem, This fulfills the promise made in the previous subsection on the Léwenheim-Skolem theorem, that second-order logic cannot be axiomatized cffectively. Elimination of the Law of Excluded Middle For Hilbert, the completed infinite already made its appearance, albeit im- plictly, in the use of instances of the Law of Excluded Middle (LEM) of the form Vn(F(n) = 0) V =Wn(F(n) = 0); these in turn lead to Wa(F(n) = 0) V 3n(F(n) # 0) and the method of proof by contradiction for sim- ple existential results. This is what made PA problematic for Hilbert and brought him to call for a proof of its consistency. In 1930, Arend Heyting set up a formal system of logic without LEM which was acceptable to the Brouwerians and has thus come to be called intuitionistic logic. One can as- sociate with any axiomatic system S based on classical logic a corresponding. system S* based on intuitionistic logic while otherwise retaining the same mathematical axioms; obviously S* is contained in S. The particular system Pat is called Heyting Arithmetic and is denoted HA. By a straightforward argument in a paper of 1933 on the relationship between classical and in- tuitionistic arithmetic, Gédel showed that PA can be translated into HA in such a way as to preserve statements of the form ¥n(F(n) = 0) with F primitive recursive. ‘To be more precise, with each statement A of the lan- guage of PA (which is the same as that of HA) is associated a statement A’ such that if PA proves A then HA proves A’. Moreover, for A of the form Yn(F(n) = 0), A" = A Godel’s translation was subsequently extended to a wide variety of systems besides PA.” Thus once more Gidel undermined Hilbert, who had stressed establish- ing the consistency of the fertium non datur (excluded thied, or LEM) in ‘the promotion of his program. Since the intuitionists said that they too reject the completed infinite, and since HA is intuitionistically acceptable, and since, finally, PA is reducible to A by Gadel’s result, what more would a finitary consistency proof of PA accomplish in the way of eliminating the “actual” infinite? Hilbert himself gave no answer to that question. ‘The Elusiveness of Cantor’s Continuum Problem With the general crumbling of Hilbert’s program, his specific aim to use it in order to solve the continuum problem came, in the end, to nought. ‘Gadel result had a precursor in a similar one by Kolmogorov for the predicate calculus. The result for arithmetic was found independently by Gentzen and Bernays in papers withdrawn from publication when Gidel’s work appeared: see Gadel (1986), 288 “See the introductory note to his paper ibidem, especially pp. 284-285,66 Foundational problems The alternatives then seemed to be either to follow Brouwer and Weyl and grant no definite meaning and no interest to the Continuum Hypothesis, or to take seriously the claims for the reality and cogency of set theory based con the platonistic vision of the set-theoretical universe. Apparently Gadel himself took this latter position early on, though he did not elaborate his view until much later, in the 1940s. One can find only a few scattered, brief indications thereof in his writings up to 1940.‘ One carly sign of the direction in which his views pointed is in footnote 48a in Gédel's 1931 paper, evidently added as an afterthought: the true reason for the incompleteness inherent in all formal sys- tems of mathematics is that the formation of ever higher types can be continued into the transfinite ..., while in any formal system at most denumerably many of them are available, For it can be shown that the undecidable propositions constructed here become decidable whenever appropriate higher types are added ‘An analogous situation prevails for the axiom system of set theary."® ‘The full extent of Gadel’s platonism only began to emerge in his paper “Russell’s mathematical logic” (1944). His attitude toward the Continuum Hypothesis was then spelled out more specifically in the article “What is Cantor's continuum problem?” (1947, and in revised and expanded form, 1964). ‘There he stated in no uncertain terms his views that Cantor's notion of cardinal number is definite and unique and that the Continuum Hypothesis CH has a determinate truth value. His own conjecture was that CH is false, because of various “implausible” consequences it has. In any case, it was Gédel’s conviction that it made sense to try to settle CH. At the same time, he acknowledged the failure thus far to come even remotely close to a solution. And finally, it was metamathematies that served to ‘explain why it was proving so difficult to arrive at an answer. Indeed, once again, Gédel himself (had already) provided the first definitive results of that character, as follows. Consistency of the Axiom of Choice and the Continuum Hypothesis In a series of brief descriptions of results, 1938-1939, and finally at length in his 1940 monograph, Godel proved the following result: (1) if ZF is consistent then it remains consistent when we add to it the Axiom of Choice AC and the Generalized Continuum Hypothesis GCH, Wa(2"= = Xq+1), as new axioms. See the discussion of Gadels philosophy of mathematics in Feferman (1986), pp. 28 32 [reproduced in chapter 6 in this volume]. ‘del (1086), p. 181Infinity in mathematics 67 Giidel’s method of proof was to introduce a notion of “constructible set” in the language of set theory and to show that when the universe of sets js restricted to the constructible sets then all the axioms of ZF together with AC and GCH become validated. Another way Gédel had of putting this (in his system of sets and classes of his 1940 monograph), using V 10 denote the class of all sets and L to denote tho class of all constructible sets, is that (2) (i) if ZF is consistent then ZF + (V = L) is consistent, and (ii) 2F + (V =L) proves AC and GCH. Here the equation V = L expresses the assumption that all sets are con- struetible, which is shown to be true in the universe of constructible sets (a seemingly obvious proposition yet one whose precise statement requires some technical work to establish, because the relativization of the notion of constructiblity to L must be shown to be absolute, that is, unchanged thereby). ‘As we have seen, almost all the work with cardinal arithmetic requires the assumption of AC. Once its consistency with ZF is established via (2)(i), (i) it makes sense to speak of the scale of alephs Ry and then to consider the truth value of 2% = Nas1, for any and all a, But one still needs the stronger hypothesis V = L to prove GCH itself ‘What Gédel's consistency result showed was that one could not hope to disprove AC, and that if AC is assumed, one could not hope to disprove any instance of GCH using Zermelo-Fraenkel Set Theory. It was still conceivable that ZF could prove AC, or that ZEC (= ZF + AC) could prove some instance of GCH, in particular, CH itself, Godel himself made efforts to establish these results, with only partial success, and only concerning the independence of AC. In fact, no progress on this problem was made in lover twenty years, despite the general expectation that both AC and CH ‘would be independent, respectively, of ZF and ZFC." Independence of the Axiom of Choice and the Continuum Hypothesis In 1963, Paul Cohen obtained the following results: (3)(2) if ZF is consistent then AC is independent of ZF; that is, it cannot be derived from ZF; (i) if ZF is consistent then CH is independent of ZFC; (iii) if ZF is consistent then V = L is independent of ZFC + GOH. Cohen's method of proof* involved a novel technique for building mod- els of set theory, called the method of forcing and generic sets. Where "According to Gadel's views im bis article of 1947 and 1964, AC fs true and CH is false, so he would certainly expect independence of CH from ZEC 515ee Cohen (1968) for an exposition68 Foundational problems Godel had restricted a presumed model of set theory to obtain that of the constructible sets, Cohen extended such a model by adjunction of (a vari- ety of ) “generic” sets. For example, by adjoining sufficiently many generic subsets of N, he was able to construct a model of ZFC in which 2*° = No, thus contradicting CH. Subsequently, building on Gohen’s work, it was shown by Easton that, for each regular Xa, the power 2*» “can be anything it ought to be”; that is, one can arrange the simultaneous equations 2** = Nya) in a suitable model of ZF, for any function F(a) = 3, from ordinals to ordinals satisfying a few simple restrictions Consistency and Independence of Definable Well-Orderings Géder’s consistency result for (V = L) and proof that 2" =X, holds under that assumption showed (4) it is consistent to assume with ZFC that there is a definable well- ordering of the continuum R; in fact, this is provable in ZF + (V = 1), Behind (4) lies the definition of L as LJ La, where at each stage, La consists of the sets explicitly definable (in a predicative way) from the sequence of Lg for a) etc., are in fact all true, and for each ‘of these $ the consistency statement will be unprovable in S. Moreover, these statements Cons go, in strength, far beyond those of the piddling ex- tensions produced above by iterating formal reflection principles or truth definitions. So, according to this line, very strong set-theoretical principles are needed to decide statements of finitary character, Bat isn’t this argument begging the question? Sure, whatever leads fone to accept “correctness” of such § will, with one little additional step, lead one to accept the consistency statement for S. But, as Brouwer re- peatedly (and Gddel himself) emphasized, consistency does not by itself ensure correctness. While consistency of such systems may be plausible, the plausibility of such can (as I indicated at the end of chapter 2) rest on quite different grounds than the assumption of an underlying set-theoretical reality in its platonistic sense. In other words, unless one is already a set-234 Countably reducible mathematics theoretical platonist, Gadel's doctrine does not by itself provide compelling reasons to embrace the Cantorian transfinte.! There remains the question: What, if any, relevance do the incomplete- ness phenomena have to finite combinatorial mathematics in the everyday sense of the word? ‘There are a number of specific results in recent years Which it is claimed establish that relevance. We turn next to a review of those results Undecidable Diophantine Problems The results here, and the question of their relevance to everyday mathemat- ics just raised, are comparatively easy to discuss. Diophantine equations are those of the form p(z1,.-. ,2) = g(21,... .2m) where p,q are polynomials in one or more variables with coefficients in Z (the integers). A Diophan- tine problem concerns whether there exists an integer solution of such an equation, that is, whether (371,... 52 € Z)ptiy..+ 42m) = a(21y-+ 52a) Hilbert’s Tenth Problem in his 1900 list called for an effective method to determine whether or not a Diophantine equation is solvable, that is, has any integer solutions [see chapter 1 in this volume for an introduction to Hilbert’s Tenth Problem). There is a closely related group of problems for solutions in N, and for solutions in Q. Specific such questions were first considered by the Greek mathematician Diophantus (third century 4.0.); after a long lapse, the subject was revived by Pierre de Fermat in the sev- centeenth century and has been a staple of number theory ever since. Many specific Diophantine problems have so far resisted attack.” {Also it follows from the work described below that the (still unproved) Goldbach Con- jecture, according to which every even integer greater than two is a sum of two primes, reduces to a Diophantine problem. Tt happens that Godel froquently referred to his undecidable propositions as being of “Goldbach type. ‘The first step toward a negative solution of Hilbert’s Tenth Problem was made by Gédel in his 1931 incompleteness paper. He showed there that every primitive recursive function F(z) is arithmetically definable, and hence can be put in the forma Q) Plt) = 9 Qr24 -. QninR(t%y 2y-+ zn) The viewpoint ere concerning independence results for arithmetical statements shouldbe compared with that of Isaacson (1987). He argues that the truths expressible in the firet-order language of arithmetic which are not derivable in PA “contain essentially higher-order, orinfinitary concepts." This is opposite to the postion taken here, that one ray beled to accept such results without auatening higher-type notions in the platonist senae, However, a regards the finite combinatorial independence reauts to be discussed in the next subsection, Isaacson and I agree that they are obtained by a transmutation ‘of essentially metamathematical statements coded in the language of arithmetic “Tat the time I wrote the article from which ths chapter was drawn, the example given ‘of the most famous such problem was Fermat's Last Theorem. This was subsequently settled by A. Wiles by very advanced methods; soe chapter 1 in this volume,Is Cantor necessary? 239 where each Q, is the universal (¥) or existential quantifier (3), the vari ables 2, range over N, and Ris built up by A,V, and - from polynomial equations. It follows that every II statement Vr F(z) = 0 is likewise defin- able in that form. Gédel concluded from (1) that for each («)consistent formal system S containing a sufficient amount of number theory, there are arithmetical propositions A which cannot be decided by S, that is, such that neither $F A nor $F ~A. In the later development of the theory of effective computability at the hands of Church, Kleene, Turing, Post, and others, attention shifted from individual problems A undecidable relative to given formal systems S, to effectively undecidable problems for subsets C of N, that is, for which there is no effective method to determine, given an arbitrary x in N, whether or not 2 € C. By Church's thesis, the effectively decidable sets C are exactly those which have a general recursive characteristic function, and then these are exactly the same as the A@ sets. Church gave examples of recursively enumerable sets which are not recursive, in other words, which are in the class £9 but not in Tf Beginning in the 1950s, considerable progress was made on Hilbert’s ‘Tenth Problem through the work of M. Davis, I. Putnam, and J. Robinson. ‘They succeeded in showing that if variable exponents are permitted, every Ef set C is definable in the form (2) BEC me Bays... enlplty21y--- 52m) = GF 21y--+ znd where p,q are exponential integer polynomials. Finally, in 1970, J. Matiya- sevich succeeded in establishing a result of the same form for onfinary inte- ser polynomials; this finally showed the answer to Hilbert’s Tenth Problem to be negative. Subsequently, Matiyasevich and Robinson succeeded in showing that it suffices to take nm < 13 to represent as in (2) every recur- sively enumerable set with ordinary p,q. This and a number of other results| concerning Hilbert’s Tenth Problem are reviewed in Davis, Matiyasevich, and Robinson (1976). A few years later, Matiyasevich managed to reduce the number of variables in the representation (2) of recursively enumerable sets from 13 to 9 for ordinary polynomials, and even further to 3 variables {for exponential polynomials.* Now, returning to undecidable propositions, it follows by Gédel's re- sults that for suitable consistent systems S of arithmetic, there are true II? statements A of the form (3) AS Va, 245.06 talPl@, 21y..- 52m) FO, Zy-- s2n)} which are not provable in S; by formalizing the work described above, this ccan be reduced to n <9 for ordinary integer polynomials p, 9 Thor references to this further work sew Mathematical Reviews, 81f; 09055 [and Matiyesevich (2999),236 Countably reducible mathematics Impressive as these results are, the problems they concern are still very distant from the bread-and-butter Diophantine problems of everyday num- ber theory (“everyday” over the last three hundred years), because the number of variables is so much larger than is considered in such problems, and the complexities of the polynomials involved are so great (according to various measures of complexity). ‘There is no evidence that these un- decidability and incompleteness results have any relevance to the classic unsettled problems that have challenged generations of number theorists [Besides the example of the Goldbach Conjecture mentioned above, it has been shown by Davis, Matiyasevich, and Robinson (1976) that the Riemann Hypothesis —which is one of the most important outstanding problems in number theory—reduces to a Diophantine problem. However, the numbers of variables and degrees of the polynomials required for this might be rather large.] Though the chain from the original undecidable propositions to the tundecidable propositions of the form (3) is a long and technically compli- cated one (so that the point of departure recedes into the background), it remains the case that the problems thus shown to be undecidable were not of any prior mathematical interest—they have simply been derived in fa step-by-step process from problems “cooked up” to demonstrate the in- completeness of formal systems. As a further reinforcement to the view here that the results on undecidable Diophantine problems are irrelevant, to everyday mathematical concerns is that the character of these problems is insensitive to the formal system considered. If one takes A in (3) to bbe equivalent to Conpa, there is no way to tell it apart from an A taken equivalent to Conzec oF from one which is equivalent to the consistency (with ZFC) of the existence of measurable cardinals. This suggests that the mathematical content of the resulting individual Diophantine prob- lems simply has nothing to do with the internal mathematical content of the formal systems from which their independence is established. How much different. it would be if one showed that some outstand- ing number-theoretic TI? statement A {such as the Goldbach Conjecture or the Riemann , Hypothesis] is not provable in PA, or even more stril ingly, in ZPC—and thus demonstrated why it's s0 difficult to prove it (if true)! There is nothing in the work on undecidable diophantine problems to suggest that one is anywhere near obtaining such a result—or anything comparable—nor that further efforts in chopping n down in (2) or (3) is going to get one any closer to achieving such a result. The situation at present is thus analogous to that concerning transcendental numbers with which we began this story in chapter 2; Liouville and Cantor showed how to construct (explicitly or implicitly) transcendental numbers, but mathe: 1aticians wanted to know whether the numbers w and e are transcendental. For those results the work of Liouville and Cantor was useless. All they did was demonstrate that the effort to show specific numbers to be transcen- dental numbers was not a waste of time, since such numbers do after all exist. Working number theorists want to know about the truth of specific nurmber-theoretic problems which can be put in diophantine form. It isIs Cantor necessary? 237 highly unlikely that one will simultaneously demonstrate the unprovability of, say, the Riemann Hypothesis RH from certain S and the truth of RH, ‘as was the case with the nonprovable propositions produced by Gadel, But mathematicians would no doubt sit up and take notice if merely unprovabil- ity were established. All that the chain of work from Gédel to Matiyasevich permits one to say currently is that efforts to obtain such independence re- sults for propositions of prior mathematical interest are not necessarily a ‘waste of time. Independence Results for Statements of Finite Combinatorial Character Starting in the mid 1970s, a number of interesting independence results have been obtained with respect to a wide range of formal systems for state- ments which are prima facie relevant to finite combinatorial mathematics in its everyday sense. The expository article Simpson (1987a) provides an ‘excellent introduction to this atea of work, and has been quite useful to me in the following, Many of the results are “finitizations” Pyix of strong, in- finitary propositions P, where P implies Pr. In some cases we even have P equivalent to Pyan. The classic example of this sort is Kénig’s Infinity Lemma, according to which if T is a finitely branching infinite tree then T has an infinite branch and (as is obvious) conversely. For T represented as a collection of finite sequences in N, the statement P that T has an infinite branch is ©}, while for T finitely branching the equivalent statement Pyin that it contains infinitely many nodes is TI9, and even IQ for branching with fixed bounds. ‘As Simpson points ont, Konig in his 1927 paper “Uber eine Schlussweise ‘aus dem Endlichen ins Unendliche” thought of his lemma as a method of obtaining infinitary results from finitary ones. Indeed, one of the most striking early applications of KL (Kénig’s Lemma) was Godel’s use of it in his proof of the completeness theorem for first order predicate logic: if a sentence A is consistent in that logic then it has a model.® ‘The hypothesis, of consistency is II?, but this is used to build a binary branching infinite tree whose nodes correspond to sequences of statements or their negations consistent with 5. In this ease we obtain a 3} conclusion from a II? hypoth- esis. The argument requires classical logic and can be carried out formally within PA, since the model constructed is arithmetically definable. (This situation is in a way a vindication of Hilbert’s view discussed in chapter 2 that the completed infinite is already implicit in the use of the Law of Excluded Middle when combined with quantifier logic.) Now, to continue with Simpson's point, the newer results discussed here proceed in the opposite direction. Starting with P which is independent, NGidel did not explicitly acknowledge the use of Kéniq's Lemme; ef. the discsasion in Gédel (1986), pp. 53-54, of his 1929-1990 work on completeness,238 Countably reducible mathematics of a formal system S, the effort is to obtain a finitary consequence Pin of P which is still independent of S. The first striking example of such was provided by the theorem of Paris and Harrington (1977), which is a modified form of the famous combinatorial theorem due to F. P. Ramsey in 1930, Ramsey's theorem concern partitions of the set of all k-element subsets of a set X, that is, a X= Cu..UG where [X]t = {¥ :¥ © X and card(¥) = k} and Ci,... ,C¢ are pairwise disjoint. We deal here only with denumerable X and may assume X CN. ‘The Infinite Ramsey Theorem P is the statement that if X is infinite then for any F and partition of [X}* as in (1), there exists an i < € and an infinite subset of ¥ of X such that 2) wicca. In the case k = ¢ = 2 this is interpreted as saying that any infinite graph whose edges (that is, members of [X]?) are colored by one of two colors (C; or C2), there exists an infinite subgraph all of whose edges have the same color. Ramsey proved a finitary form Prin of P which is as follows: (3) for any &, €, and m there exists n such that for any X with card(X) ='n and partition [X]* = CiU...UCz, there exists, i<€and ¥ CX such that card(¥) > m and (¥]* € Cy This statement can be seen to be a consequence of P by use of KL; if (3) is assumed false, then a finitely branching tree T can be constructed such that an infinite branch X through it violates the conclusion of the Infinite Ramsey Theorem. It turns out that this Finite Ramsey Theorem (3) can be proved in Peano Arithmetic PA. The surprising result found by Paris and Harrington is that a slightly modified form P7,, of (3) is independent of PA. This Modified Finite Ramsey Theorem differs from (3) only by addition, to the conclusion, of the requirement that card(Y) > min(¥'), where min(Y) is the least element of ¥. The argument that P implies Pj, can again be carried out by contradiction and use of KL (though, as pointed out by ‘Simpson, this use of KL is inessential). Since the work of Paris and Harrington, a number of other results of finite combinatory character have been shown to be independent of PA. One of the simplest is a theorem due to Goodstein concerning the effect of shifting bases in the representations of natural numbers to various bases b, which has been shown by Kirby and Paris to be independent of PA.* It should be emphasized at this point that the statements Pyin thus shown independent of PA are recognized to be true by infinitary methods See Simpson (1987a) for detailsIs Cantor necessary? 239 which go beyond PA. They can in fact be proved in the second-order system (II2,-Ca) based on the arithmetical comprehension axiom with full induction. Moreover, Pjin is equivalent to the I-consistency of PA (a notion intermediate between w-consistency and consistency}. Incidentally {as is discussed in the next and final section), the system (I19,-CAY| with restricted induction proves the same arithmetical statements as PA.° ‘The system (I12,-CA) is predicatively justified, since the second-order variables can be interpreted as ranging over the arithmetically definable subsets of N. Schiitte and I both analyzed in formal terms the transfinite iteration of the formal process of introducing sets by definitions referring only to previously determined collections of sets, where the iteration ex- tends only to those ordinals corresponding to previously recognized well orderings. The limit of these is a certain countable ordinal To.” This char- acterization of predicativity yields a sequence of ramified systems of analysis Re each of which is predicatively justified for a . Jin my 1964 paper and in several subsequent papers I produced a variety of single unramified systems S which are proof-theoretically of the same strength as U Rala < L] and hence are locally predicative.? Another such system of strength Ty has been introduced and utilized by Harvey Fried- man, and denoted ATR by him; in our notation this is denoted ATRY, since it uses restricted induction on N. The locally predicative system (ATR}) is itself relatively weak subsystem of the patently impredicative system (ii-cay. ‘The point of all this here is that Friedman found a finite combinato- rial version Pyin of an existing infinite combinatorial theorem P known as Kruskal’s Theorem, such that Prin is independent of ATR} and hence can- not be proved by predicative methods. In this case Kruskal’s proposition P concerns a certain (relatively simple) embeddability relation < between finite trees; it says that the collection of such trees is “woll-quasi-ordered” under <, that is, (4) for any infinite sequence T;,Ta,-.. ,Tm,.-- of finite trees there exist i,j with i < j and T; < Ty (In other words, there is no infinite descending sequence in this collection under the relation <* defined by T <* T’ -» T’ ¢ T). Note that (4) is ‘equivalent to a TI}-statement. Now the following is Friedman's finite form Prin of Kruskal’s theorem: (5) For any & there exists m such that for any finite sequence Ty.Ta,--. 5 Tn of trees with card(T;) < ki for all t 2.) At any rate, what the above reductions demonstrate is that the classical systems in question have an alternative constructive justification which does not require anything like belief in a preexisting totality of subsets of N. Moreover, the same kind of alternative constructive justification has been given for far stronger theo- ries than (II}-CA) + BI, namely, for the system (4}-CA) in the work of Buchholz, Feferman, Poblers, and Sieg cited above, and for (A$-CA) + BI in further work of Jiger and Poblers.* At present, none of the indepen- dence results for finite combinatorial systems of the kind described above come close to exhausting the strength of systems for which we thus have a completely opposite (nonplatonist) constructive justification, In the preceding discussion I have, for the sake of argument, taken for granted that the independent statements produced are indeed relevant to everyday combinatorial mathematics, even though they do not soive problems of prior mathematical interest. But even that may be challenged it seems to me that as one passes from statements of the sort provided by Paris-Harrington and Kirby-Paris for PA, through those provided by Friedman for ATR} and (I1}-CA)}, on to those provided by Buchholz for {I1}-CA) + BI, the connection to practice becomes more and more tenuous. Viewed as an observer from the sidelines, there is a nagging feeling that the statements sill have a "cooked up” look. This already appears with the use "See Buchholz et a. (1981) or a general review of thie work and further references.242 Countably reducible mathematics the Paris Harrington statement of the condition eard(Y) > min(¥), and Friedman's finite form of Kruskal's theorem of the condition card(T,) < ki, Examination of the details of the argument in Simpson (1985) for the independence of the latter from ATR{ makes clear how the imposition of the linear condition on growth allows one to pass from that statement (with the aid of Kénig's Lemma) to the well-foundedness with respect to a wide-enough class of (possible) descending sequences to carry through & proof-theoretical demonstration of the I-consistency of ATR{. Note that the Paris-Harrington statement, which is most clearly relevant to results| of established mathematical interest (finitary forms of Ramsey's theorem), does nothing to support Gédel's doctrine, since it can be proved in a mild, predicatively justified, extension of PA. But, in my view, the further one moves to support Gédel's doctrine by independence results for increasingly stronger systems, the less convincing is the case for the relevance of such? This section would not be complete without: some discussion of the leap that has been taken in Friedman (1986). ‘The effort there is to pro- duce finite combinatorial statements Prin independent of relatively strong systems of set theory, and in particular of ZFC + {“(there exists an n- Mahlo cardinal)” Jncw. While the statement thus produced"? does make use only of concepts from ordinary finitary mathematics, it is extremely remote from any ordinary mathematical reading. (Indeed, so much so that it would be unrewarding to reproduce the statement here—-the reader must see for himself in Friedman's paper.) Here, I am evidently in complete dis- agreement with Friedman that his statement is “readily understandable.” But Friedman himself goes on to say that this statement “is still not quite as simple or as natural as we would like [and] we are continuing to strive for an improved form."* What Friedman believes his result in this instance accomplishes is to “open up, for the first time, the realistic possibility, if not probability, that strong abstract set theory will prove to play an es- sential role in a variety of more standard finite mathematical contexts.” Tf that case can be made then, according to Friedman, “this would open up a foundational crisis of nearly unprecedented magnitude since we seem to have no way of convincing ourselves of the correctness or consistency ‘of such set theoretic principles short of faith in our very uneasy intuition about them." In other words, Friedman is hesitatingly plumping for some form of Gadel's doctrine. “jn personal communications to me, both Friedman and Simpson have taken issue with the views expressed inthis paragraph, particulary with respect to the comparison of the finite form of Kruskal's theorem given in the text and the Paris Harrington statement (PH). Lagree with them that the former statement is more readily understandable (oF Visualizable) than the latter, and also that cardinality restrictions on the sizes of finite trees are more natural than the ad hoe cardinality condition appearing in PH. However, this does not affect my overall conclusion atthe end ofthe paragraph, which also concerns ich stronger statements "0Fyiedaan (1986), proposition VIL “[His most recent efforts in that direction are presented in Friedmen (1998),) WPhe quotations are fom Friedman (1986), pp. 92-93,Is Cantor necessary? 243 My general arguments against that doctrine in the last section, except one, apply equally to this specific case. Once more there is a petitio prin- cipit: would we believe Friedman's independent finitary statement. to be true if we did not believe in the correctness of ZFC + Vn (“there exists an n-Mahlo cardinal”) in some independent set-theoretical reality? So how docs it justify the latter to demonstrate independence of the former from a slightly weaker system? Since the statement is clearly fabricated to do 1 job, and the evidence for its truth begs the question, the necessary use of such strong systems for everyday finitary mathematics has yet to be demonstrated. As I read Friedman's statements above, he agrees that the cease has yet to be clinched; clearly, we differ on how much farther one will have to go in order to do that decisively, if at all. But my overall conclu sion (announced in the introduction to chapter 2) about the independence results discussed in this section is even stronger: the case remains to be established that any use of the Cantorian transfinite beyond No is necessary for the mathematics of the finite in the everyday sense of the word." Countable Foundations for Applicable Mathematics In this final section,** I shall explain the basis for the second claim an- nounced in the introduction to chapter 2, that higher set theory is dispens- able in scientifically applicable mathematics. The argument for this follows the trail blazed by Weyl in his 1918 monograph Das Kontinwum, but goes much farther by making use of modern logical tools. First of all, the work sketched by Weyl can be formalized in (I1°,-CA)>, which is a conservative extension of PA; that is, for any first order A, if (TI,-CA)f F A then PA A. This conservation result can be proved very simply by use of Gédel's completeness theorem, as follows. If PA lf A then there is a model M of PA +{74}; one then expands M to a model M" of (II2,-CA)t + (74) by taking the range of the set variables to be the collection of all first- order definable subsets of M. There are also standard proof-theoretical techniques which may be used to prove this conservation result by strictly finitary means. Thus, the portion of classical mathematical analysis that can be formalized in (I19,-CA)I, following Weyl’s plan, rests on first-order Peano Arithmetic as a foundation. Since the general notion of real number is defined in the wider system, the conservation result shows that such uses of the uncountable in classical analysis can be eliminated. ‘Though Weyl did not himself carry his plan very far, the direction he set for it was clear enough to see how to formalize substantially all of the classical analysis of piecewise continuous functions of a real variable devel- oped through the nineteenth century (and relatedly for complex analysis). =[[Por a more recent discussion of the issues in this and the preceding sections, sce Feferman (098) | “(The material of thi section overlaps that of chapters 1 and Lin this volume, vwhowe writing it preceded |24 Countably reducible mathematics Since this part of analysis already inchides a considerable portion of scien- tifically applicable mathematics, the execution of Weyl's program would go far toward substantiating the claim made above. However, one would also have to be able to provide a countable foundation for significant portions of ‘twentieth-century analysis and accessory parts of mathematics in algebra and topology in order to establish the claim in full. Among other things ‘one would want to be able to deal with the Letesgue integral and more gen- eral theories of integration, as well as with a variety of function spaces and more general abstract spaces (Banach spaces, Hilbert spaces, etc.), and the operators on them, which form the subject matter of functional analysis. One way in which this might be done is to deal only with countably pre- sented mathematical objects (functions, sets, structures, spaces), each of which is determined by a countable amount of information and hence can be represented by a subset of N. While this approach is possible in prin- Ciple, it is not very natural. It is preferable to find a system S for which fone can proceed in a direct way from a body of mathematical notions and theorems to the formalization of such in the system, and such that $ can be given a countable foundation (as (I12,-CA)I is justified by PA). However, a system of this type would have to incorporate higher type notions. For, as soon as we wish to speak about functions of real numbers in a direct way in general, we must move to third-order concepts; then direct discussion of functionals (such as vatious linear operators on functions) would require dealing with (prima facie) fourth-order concepts, and so on. Thus the aim would be to find a finite type theory $ which can be given a countable foundation, say by proof-theoretical reduetion to PA, and in which scientifically applicable twentieth-century mathematics can be di- rectly formalized. The first part of this aim was realized in two ways by systems presented in Feferman (1977) and Takeuti (1978). But these sys- tems are still not as convenient as one would like for the second part of the aim, The reason is that types are fized syntactic objects in these systems, 0 there is no simple direct means to handle subtypes given by separating with respect to properties having variable parameters. A second reason for wanting variable types is that, following modern abstract mathematics, cone wants to speak of arbitrary structures of any given algebraic, topolog- ical, or analytic kind and there is no natural sense in which these should be restricted to range over a fixed type. For example, one would like to speak of all groups (or all linear spaces, or all Banach spaces, etc.) and not just of all groups (etc.) within a given type. What is thus required is a formal system in which the types themselves may be variable and in which subtypes can be freely separated by suitable definitions. Such a sys- tem would approach set theory in appearance, Indeed, Friedman (1980) provided a theory ALPO which is contained in ZFC and is conceptually rich, yet is a conservative extension of PA. However, Friedman's system is semi-intuitionistic (in the sense that classical logic is applied in it only to ‘number-theoretie equations) Since the use of classical logic pervades modem mathematics, in my view it is preferable to have a formal system with variable types using clas-Is Cantor necessary? 45 sical logic in full and having a countable foundation, say by reduction to PA. Such a theory has been described in Feferman (1985a)"®; it was there denoted Res-VT + (2) but is denoted VT, in the following (“VT" abbre- viates “Variable Types" and “y” is for the unbounded minimum operator described below), While the formalism employs set variables, it differs from set theory in that the notion of function is treated as a primitive and not reduced to the notion of set. This conceptual separation is characteristic of| certain approaches to constructive mathematics, At the same time it ac- cords with the highly asymmetric use that working mathematicians make of functions and sets (in general, in practice, functions are not nearly as complicated as sets). It is not possible here to go into details of the system VT,t, but the following sketches its main features. Besides class variables (or “type vari- ables”) X,¥,Z.... there are class terms U,V,W, ... built up from class constants and variables by the following operations: (i) if U,V are class terms then so also are U x V and (U —» V); and (ii) ifU is a class term and Ais a bounded formula (defined below), then {1 € U : A} is a class term, Individual terms are all introduced in context: for any class term U (which ‘may contain variables) there is a list of individual variables 2", y¥,2",... of type U; farther terms are built from these and individual constants b pairing and projections (corresponding to Ux V), and by function ap plication and abstraction (corresponding to U —+ V). Thus we have the full means of a typed A-calcutus but with variable types. Unlike ordinary type theory, the formalism permits equations s = ¢ between terms of arbi- trary type. Formulas are built from atomic formulas (s = t) by ~4A,V,—, and universal and existential quantification with respect to both class vari- ables and (relative to any class term U), individual variables 32”(...) and vel(...)."¥ Bounded formulas are those containing no quantified class variables. (The preceding description of the syntax requires a simultane- cous definition of class terms, individual terms and formulas). One defines (t€ U) as 324 (t = 2) and {2 EU : A} = {29 : A},(Gz € U)A = 30, and (Wz € U)A = Wat. The basic axioms of V'T,,[ are the usual ones for abstraction and applica~ tion, pairing and projections, and separation. The further axioms concern N (a constant symbol for the set of natural numbers). This comes equipped with constants for 0 and the successor function sc(z) = 2’, and a primitive recursor ry which gives for each a € N and g € Nx N — N a function f = rw(a,g) € (N — N) with f(0) = aA Yn € N[f(n') = a(n, f(n))]. From this we can derive primitive recursion with parameters drawn from "This paper was presented to a meting in Bogoté, Colombia, im 1981. ‘The excep- tional delay in publication of the proceedings of the meeting accounts for ite date. [Aa improved system W for this purpose is described in chapters 13 and 14] “3The system in Faferman (1985) did not employ quantified class variables, using sul only for pure statements of generality, ‘The present extension is conservative over (hat system246 Countably reducible mathematics any classes. The axiom of induction for N is given in a form restricted to classes defined by characteristic functions: (1) f(0) =OA Vr € N(f(x) = 0 — f(z") = 0) + Wz € N( f(z) = 0). Finally, there is an axiom for a function gi € (N + N) -+.N with (2) ulf) = migK (2) = 0] i (x € N)[F(z) =O), otherwise 0; wis called the unbounded minimum operator. Using it we can define an operator Ex with Ex(f) = 0 ++ (3x € N)(f(x) = 0), ‘The main mathematical result obtained for this system is that @) VT,f is a conservative extension of PA. ‘The proof of (3) described in Feferman (1985a) proceeds by a series of reduction steps—first from variable types to constant types, then by elim- ination of subtypes, and finally by reduction to the system Res-Z» + (1) of, Feferman (1977); it was there shown how to reduce the latter system to PA by proof-theoretical means. All of this argument is finitary; furthermore, conservation with respect to arithmetical statements is preserved at each step.” While (3) shows that VT,,{ has a countable foundation, many sets which are ordinarily thought of as being uncountable can be defined to exist in the system. To prepare the way, Z is defined as usual in terms of Nx N, and Q in terms of Z x (Z-{0}}). Then the class R of real numbers is defined to be the class of all Cauchy sequences of rationals, that is, all functions from N to Q satisfying the Cauchy convergence criterion; an equality relation =n is introduced to tell when these represent the same real number. The class Fung of al real functions is then defined to be the subclass of (R + R) consisting of all f which preserve =n. We can then proceed to define the ‘lass ofall functionals on these and so on, Of course the complex numbers C are defined as usual from R via RXR. Since n-tupling can be explained in terms of pairing, one has all finite-dimensional spaces R” and C"; one also has Cantor space and Baire space via N ~ {0, 1} and N + N, respectively. All of these spaces may be shown to be locally sequentially cornpact; that is, every bounded sequence contains a convergent subsequence. (The proof uses Kénig’s Lemma, which in turn is proved by a combination of primitive recursion and the Bw operator.) The Cauchy completeness of these spaces follows. More generally one deals with arbitrary complete separable metric spaces, for which fundamental results like the Baire Category Theorem and the Contraction Mapping Theorem (existence of a fixed point) can be proved. ‘The space Ras dealt with in VI,f cannot be proved complete in the sense that the Lu.b. (or g..b.) axioms holds for arbitrary sets of reals; one The eduction to PA of the improved system W refered to in the postecript to fn. 12 above is carried out in Feferman and Jager (199%) and (1996)Is Cantor necessary? 247 only has that for arbitrary sequences of reals. Since the Lu.b. (or gb.) axiom for sets of reals is applied ubiquitously in analysis, one must in each case see whether such an application can be replaced by use of the corresponding principle for sequences. Often that is very easy but in some ‘cases it requires more care; finally, there are certain cases in actual analysis where this replacement simply catinot be made. One such is when Lebesgue measure meas(X) of a set X C Ris defined in terms of its outer measure ‘meas*(X), which is the gb. of {meas(¥) : X ¢ ¥ and ¥ is open) (where for open ¥, meas(Y) is the sum of the lengths of the components of ¥) Failing this, we cannot deal with nonmeasurable sets. But for applications it is only necessary to deal with measurable sets X, and for these a direct, definition of measure is possible in terms of sequential open covers of X and of the complement of X. Similarly measurable functions can be dealt with directly in terms of limits (a..) of sequences of step functions. By this means we can represent the positive theory of Lebesgue measurable sets and functions (and similarly for more general theories of measure and integration). Finally, one can develop substantial portions of functional analysis in VT, { for linear operators on separable Banach and Hilbert spaces, in- cluding the principal results for the spectral theory of compact self-adjoint operators. I have verified that usable forms of the Riesz Representation Theorem, Hahn-Banach Theorem, Uniform Boundedness Theorem, and the Open Mapping Theorem can all be proved in VT, |. It seems, then, that all of applicable classical and modern analysis can be developed in this, conservative extension of PA. ‘There is further detailed evidence which can be brought from another corer to support this claim, Namely, in his famous 1967 work, Errett Bishop gave a new constructive redevelopment of classical and modern analysis which he carried out in great detail for a number of fundamental results (many of which could be considered test cases). Bishop said that each of his theorems provides a constructive substitute A* for a classical theorem A. The relationship is that A implies A* and that A* implies A by use of classical logic. In fact, according to Bishop the use of classical logic can be reduced in each case to that of what he called the Limited Principle of Omniscience (LPO), namely, that (4) Yn € N(f(n) = 0) v 3n € N(f(n) #0) holds for each f € (N — N).!* Now Friedman had found in 1977 a sub- system of intuitionistic ZF in which all of Bishop's constructive analysis (with a later form of his theory of measure) can be formalized, and which is conservative over HA. In my paper “Constructive theories of functions and classes"™® I provided an alternative constructive theory of functions “This explains Friedman's use, in his 1980 paper, of “ALPO" as an abbreviation for "Analysis with the Limited Principe of Ommniscienca” 1SFeterman (1979)248. Countably reducible mathematics and classes Ta, containing a subtheory EMof in which all of this work of Bishop can be formalized. I showed that EMo with classical logic is conser- vative over PA and Beeson showed its conservation over HA in intuitionistic logic.!® ‘The crucial point for the present purposes is that when classical logic is added, one regains from Bishop's work all the classical theorems A for which he found constructive substitutes A*, in a theory conservative over PA. (The system EMg is related in certain ways to VT}, without the operator; addition of the latter builds in a strong form of LPO.) Actually as far as the claim here is concerned, the detour via Bishop's work takes one through unnecessary technical complications, since those are necessitated only by his insistence on constructivity, But at least the foregoing provides another way of giving massive detailed support for the main claim of this section Still farther evidence for that has been gathered in the program of “Roverse Mathematics” initiated by Friedman and pursued by Simpson and others, which shows that even much weaker theories, conservative over PRA (Primitive Recursive Arithmetic), suffice for the development of significant portions of analysis and algebra.” Nothing here is meant to suggest that a system $ which makes prima facie use of uncountable sets of various kinds must be reduced to PA in order to show that higher set theory is eliminable in the applications of S. Any predicatively reducible system is a candidate for that and even, from a certain point of view (suggested above), are certain (prima facie) impredicative theories of iterated inductive definitions, Moreover, there is no question that in the current practice of mathematics there are & number of results in which the use of transfinite set theory is essential. While these results are far from the applications of mathematics, further efforts should be made to see whether any of them might lead to a crucial test case of the claim made here. Independently of such detailed work which puts into question the ne- cessity of higher set theory for everyday mathematics,* I am convinced that the platonism which underlies Cantorian set theory is utterly unsat- isfactory as a philosophy of our subject, despite the apparent coherence of current set-theoretical conceptions and methods, To echo Weyl, platonism. 4s the medieval metaphysics of mathematics; surely we can do better. 3 Feterman (1979), pp. 218-219. See Simpson (19868), for an introduction and references [and, now, Simpson (1998) “ln this expect, cf also the following chapters 13 and 14],